Software Costs
AI Pilot to Production: Why Contractor Software Projects Stall
Why an AI pilot to production move stalls at a service business: eight failure modes, a ten-gate readiness check, the reliability math and what it costs.
In this article
- What the data says about pilots that never ship
- Why an AI pilot to production move stalls: eight failure modes
- The readiness check: ten gates to clear before cutover
- Why your pilot's accuracy number is lying to you
- How to choose an autonomy level you can actually govern
- What will change under you: platforms, prices and retirements
- What does it cost to get a stalled project live?
- A 60-day plan to unstick a stalled project
- When not to take a pilot live
- Next steps
- Frequently asked questions
- Sources
Most projects that try to go from AI pilot to production at a service business do not fail because the software is not good enough. They stall because nobody settled three things before the demo: who owns the result, whether it can run on the messy data the business really has, and what number has to be true for it to go live. The same is true of any new system a contractor tries to roll out, whether it is an automated invoicing flow, a call-answering assistant or a replacement for the field service platform.
This article is the practical version of that argument, written for the owner or general manager of a company with a dispatcher, an office team and a fleet of trucks. It covers what the credible data says about pilots that never ship, eight failure modes that stop projects at a contractor in roughly the order they bite, a ten-gate readiness check you can run this week, the arithmetic that explains why a "95% accurate" automation is unusable across twenty steps, how the software underneath you changes without asking, and honest cost ranges for getting a stalled project live.
It is also candid about when to stop. A pilot that never goes live is sometimes the correct outcome, and the cheapest way to learn that is before you have signed a second invoice.
Key takeaway: The distance between a pilot and a working system is measured in ownership, real data and a decent plan for your busy season, not in how clever the software is. Budget for those three things first.
What the data says about pilots that never ship
Start by distrusting the headline number. The most quoted statistic in this field, the claim that 95% of generative AI pilots produce no measurable return, comes from a 2025 MIT Media Lab NANDA report. As of October 2026 the URL that report was published at redirected to a project overview page that does not list it, so this article does not rely on it. A statistic whose primary source has quietly disappeared is not one to put in front of your partners.
Here is what is documented and checkable.
Abandonment is real and it got worse. S&P Global Market Intelligence surveyed 1,006 midlevel and senior IT and line-of-business professionals across North America and Europe for its Voice of the Enterprise: AI & Machine Learning, Use Cases 2025 study. It found that the percentage of companies abandoning the majority of their AI initiatives before they reach production "has surged from 17% to 42% year over year," with organizations reporting that an average of 46% of projects are scrapped between proof of concept and broad adoption. These are large-company surveys; no comparable number exists for contractors, and nobody should pretend to have one.
Forecasters expect a lot of cancellations. Gartner predicted in June 2025 that "over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls". In May 2026 it added a second prediction: by 2027, 40% of enterprises will demote or decommission autonomous AI agents because of governance gaps identified only after production incidents occur.
The money question arrives before the technical one. In IBM's 2025 CEO study of 2,000 CEOs in 33 countries, respondents said only 25% of AI initiatives had delivered the expected ROI over the previous few years, and only 16% had scaled enterprise-wide (IBM). Deloitte's survey of 1,854 executives in Europe and the Middle East, published in October 2025, found most reached satisfactory ROI on a typical AI use case in two to four years, against the seven to 12 months usually expected of technology investments (Deloitte). A pilot that cannot show a measured return against a baseline will struggle to get a second payment, however well it demos. The method for building that case from your own numbers is in our guide to the payback arithmetic for replacing software.
Two things follow. The base rate for getting a pilot live and keeping it there is poor, so a cautious plan is not pessimism. And the differences between projects that ship and projects that stall are organizational and operational, visible before anyone writes a line of production code.
Why an AI pilot to production move stalls: eight failure modes
These are the eight patterns that actually stop projects at a service business, ordered roughly by when they show up.
1. Nobody agreed what "working" means
A pilot is judged by whether the demo impressed people. Going live needs a threshold. Anthropic's guidance on defining success criteria is blunt about what a usable target looks like: specific, measurable, achievable and relevant, and it notes that "most use cases need multidimensional evaluation along several success criteria."
For a contractor, that means numbers a dispatcher would recognize. Share of completed jobs invoiced the same day. Share of inbound calls answered and booked correctly. Share handled with no fix by a person. If your project has one number ("it seems accurate"), or none, you do not have a go-live decision. You have an argument waiting to happen.
2. The pilot ran on data the real business cannot give it
Pilots run on an exported spreadsheet, ten tidy jobs or a copy of last month's tickets. The live business has a customer record split into three because the name was typed differently, equipment installed in 2014 with no model number, and half of the job history in a technician's text messages. Gartner's February 2025 research found that 63% of organizations either do not have, or are unsure whether they have, the right data management practices for AI, based on a July 2024 survey of 1,203 data management leaders, and predicted that through 2026 organizations would abandon 60% of AI projects unsupported by AI-ready data.
Its April 2026 survey of infrastructure and operations leaders puts a sharper number on it: 38% of leaders who faced setbacks said poor data quality or limited data availability was a direct cause of AI project failure. The same applies to any new system: if the old data is a mess, the new tool inherits the mess on day one.
3. Nobody owns it, and the person who sponsored it is on a roof
Software with no owner drifts. Connections change, a tool updates, a rule stops matching how the business works, and nobody whose job includes noticing. In a contractor the sponsor is usually the owner, who is also the best salesperson and the person everyone calls. The owner cannot be the day-to-day owner of a project. Name the office manager or operations lead, give them hours, and write the name down.
Gartner's September 2026 prediction that 70% of enterprises will abandon agentic AI built by vendor forward-deployed engineering by 2028 is fundamentally an ownership failure: the release observes that "FDE engagements often fail structurally before they fail technically", because scope, incentives, governance, IP ownership and exit were never settled. The small-company version is the developer who builds something, gets busy and stops answering email.
4. The project started at the wrong time of year
This is the contractor-specific one, and no survey measures it. Every trade has a season in which nobody in the office has an hour to spare: the first heat wave for HVAC, the first freeze for plumbing, storm season for roofing and restoration, the spring rush for landscaping. A project that needs two hours a week from your dispatcher stops dead in that month, and a cutover scheduled for it breaks dispatch at the worst possible moment.
Plan backwards from your peak. Anything that asks the office for real time (mapping the process, testing, training) belongs in the slow months. Anything that changes how trucks are dispatched should be live at least a month before the peak or wait until after it.
5. The field app is an afterthought
Technicians are the most change-resistant users in the business and the most expensive to disrupt. They work with one bar of signal, one hand free and no patience for a form with eleven required fields. A pilot that was demonstrated on a laptop in the office has not been tested on the people who produce the data everything else depends on. If photos, notes and completion status do not arrive from the truck, the office automation downstream has nothing to work with.
The sequence that works is office modules first and the field app last, with a technician or two testing early and the rest moved over only once the office side is trusted.
6. Nobody cancelled the old subscription
The quietest failure is financial. A new system goes live, the old platform stays on because "we might still need it," the annual contract auto-renews, and a year later the business is paying for both. A project that was meant to cut the software bill has raised it. Gartner's April 2026 survey of infrastructure and operations leaders found failures concentrated where leaders expected AI to "fix long-standing operational issues", which is a reminder that the goal has to be specific.
Write the renewal dates of every tool the project touches on one page before it starts, with the cancellation notice period next to each. Decide in advance which subscription goes away and on what date, and what has to be true for that to happen. A project with no cancellation date has no savings.
7. Per-step reliability does not survive multi-step work
This one has arithmetic behind it, and it gets its own section below. The short version: an automated workflow's end-to-end success rate is roughly its per-step success rate raised to the power of the number of steps, and that curve is brutal. A June 2025 study, More Vulnerable than You Think: On the Stability of Tool-Integrated LLM Agents, found AI agents are highly susceptible to errors at every stage of tool use, and that scaling the model up does not substantially fix tool invocation reasoning.
8. Trust never transferred to the people who have to use it
The quietest failure mode and often the final one: the office manager does not believe the numbers, so she keeps her own spreadsheet. A tool that people route around is indistinguishable from a tool that does not exist. Adoption is not a launch-day activity; it is a design input. Let the people who do the work see the old and new outputs side by side, and fix what they point out before you ask them to trust it.
The readiness check: ten gates to clear before cutover
This is the checklist to decide whether a pilot is ready to be promoted. Score each gate 0 (not started), 1 (partial) or 2 (done and evidenced). Anything below 16 out of 20 means you are shipping risk, not a product. Any gate at 0 in the Control or Operate group should block cutover outright. If you have not started a pilot yet, score the idea first with our readiness checklist for replacing software; these gates pick up where it ends.
| # | Group | Gate | What "done" looks like |
|---|---|---|---|
| 1 | Decide | Named business owner | One person, not a committee, is accountable for the outcome after launch, with hours set aside for it |
| 2 | Decide | Success numbers written down | A number per measure: jobs handled with no fixes, escalations, time to invoice, cost per job |
| 3 | Decide | Rollback and kill switch | An agreed error rate that triggers going back to the old way, and a tested way to switch it off in minutes |
| 4 | Evidence | Real jobs it has never seen | At least 100 real past jobs, checked by someone who does the work, including the ugly ones |
| 5 | Evidence | Checked step by step | You know how it reached each answer, not just that the final answer looked right |
| 6 | Control | Autonomy matches measured reliability | Permission to write follows evidence, not hope: see the ladder below |
| 7 | Control | Least-privilege access and a log | Separate credentials per tool, and every action recorded with who or what approved it |
| 8 | Operate | Costs and runs visible | Every run traceable, with cost per completed job reported monthly |
| 9 | Operate | Plan for the platform changing | A named replacement for each vendor or model it depends on, and a test set you can re-run |
| 10 | Operate | A human path someone staffs | A real queue with a real owner and a response-time target, not an unmonitored inbox |
Key takeaway: Gates 1, 3 and 10 are organizational and cost nothing but a decision. They are also the three most commonly missing when a pilot stalls.
The structure mirrors how larger organizations frame the same problem. If you want a neutral reference to check it against, the NIST AI Risk Management Framework organizes the work into Govern, Map, Measure and Manage, and is voluntary. A contractor does not need to adopt it; the ten gates above are the small-company version of the same questions.
Why your pilot's accuracy number is lying to you
Here is the most useful piece of arithmetic in automation, and the reason experienced teams get nervous when someone says "it works 95% of the time."
If a workflow completes a job by taking several independent steps, and each step succeeds with the same probability, the probability that the whole job succeeds is approximately the per-step rate raised to the power of the step count. Steps are not perfectly independent in practice, and well-built systems recover from some errors, so treat this as a floor rather than a forecast. It is still the right mental model, and it explains almost every "but it worked in the demo" conversation.
| Per-step success | 5 steps | 10 steps | 20 steps | 50 steps |
|---|---|---|---|---|
| 95% | 77.4% | 59.9% | 35.8% | 7.7% |
| 99% | 95.1% | 90.4% | 81.8% | 60.5% |
| 99.9% | 99.5% | 99.0% | 98.0% | 95.1% |
A job-to-cash workflow is a good example. Read the technician's notes, match the job to the right customer and property, price it from the price book, attach the photos, draft the invoice, send it, record the payment, update the service agreement. That is eight steps before any exception. At 95% per step the chain finishes cleanly about two jobs in three; at 99% it finishes about 92 in 100. The 95% version looks fine in a demo and sends a wrong invoice every third job in production.
To keep end-to-end success at or above 95%, a 95% step rate buys you one step, a 99% rate buys you five, and a 99.9% rate buys you fifty-one.
Three practical consequences:
Shorten the chain before you improve the software. Replacing a chain of four manual hand-offs with one connection between two systems does more for reliability than any upgrade. If the real problem is that tools do not talk to each other, connecting the tools you already pay for removes steps instead of adding clever ones.
Measure the path, not just the destination. Google Cloud's agent evaluation documentation distinguishes final response evaluation from trajectory evaluation, which compares the sequence of steps the agent took against a reference sequence. A final-answer-only check will happily pass a system that reached the right invoice total by accident after three wrong lookups, and it will hurt you in production. That feature is in preview at the time of writing, so check its status before building anything on it.
Add checkpoints where the chain is long. If a workflow genuinely needs twenty steps, break it into stages with validation between them, so a failure costs one stage rather than the whole job.
How to choose an autonomy level you can actually govern
Gartner's May 2026 guidance proposes classifying automation by autonomy level and applying proportional controls to each. Its four levels are Observe (read-only), Advise (generates drafts that humans execute), Act with Approval (writes data or sends messages only after explicit human approval) and Act Autonomously (executes within guardrails, with humans reviewing exceptions).
The useful move is to tie the level you deploy at to the reliability you have actually measured, and to promote only when the evidence supports it.
| Autonomy level | Contractor example | Promote to the next level when |
|---|---|---|
| Observe | Reports on unbilled jobs and overdue invoices; nothing is sent or changed | The reports match what your bookkeeper finds by hand |
| Advise | Drafts the invoice or the follow-up text; a person sends it | Reviewers are fixing fewer than a few in a hundred and not just clicking approve |
| Act with Approval | Queues invoices and reminders; the office manager releases a batch each morning | Weeks of batches with almost nothing changed |
| Act Autonomously | Sends standard reminders and files photos on its own; exceptions go to a person | Only for narrow, high-volume, low-damage tasks, with a switch you have tested |
At the approval level, "human review is effective only if it remains a meaningful control": without a clear workflow, approvals degrade under time pressure into a false sense of safety. The common mistake is launching at Act with Approval because it feels safe, then discovering that the approval queue is the bottleneck and quietly loosening it. Decide the promotion criteria in advance and write them down.
Anything that touches money deserves the slowest promotion. A reminder text that goes out a day late is an annoyance. An invoice sent to the wrong customer is a phone call you do not want to take, and it is the kind of mistake that the OWASP list of risks for AI applications calls excessive agency: giving a system more permission than its measured reliability justifies.
What will change under you: platforms, prices and retirements
A pilot is a snapshot. A live system is a subscription to someone else's roadmap, and the platform changes on a published schedule that most teams never read. This applies to every tool your project depends on, and to the AI models inside it.
Models retire, on a clock, with no extensions. Microsoft's Foundry Models lifecycle and support policy sets a generally available model's retirement date 18 months out at launch. At 12 months the model becomes Deprecated, which blocks new customers. At 18 months it is Retired and "all inference requests return 410 Gone." Customers with active deployments get at least 60 days of notice before a GA retirement. The documentation is explicit that retirement dates are not extendable.
Anthropic's own model deprecations page commits to "at least 60 days' notice before model retirement for publicly released models." Recent history shows what that means in practice: Claude Sonnet 4.5 was deprecated on September 30, 2026 and retires on November 30, 2026, a 61-day window.
Preview versions are not production versions. Microsoft states it directly: "Preview models aren't recommended for production workloads." If your pilot was built on a preview model, that is a migration you have already committed to.
Your own vendors change too. A field service platform raises its per-seat price at renewal, retires an integration or changes an API. The practical response is not to fight this. It is to make every dependency a replaceable part:
- Keep the test set from gate 4 re-runnable on demand, so swapping a model or a vendor is a measurement rather than an argument.
- Keep vendor and model identifiers in configuration, never scattered through the system.
- Put a quarterly reminder on the calendar to check each provider's deprecation and pricing pages.
- Assume one model migration per year per workflow and put it in the run budget.
This is also the strongest argument for having the work operated by someone whose job it is to watch it. In Fleurant AI's Operations Stack, hosting, monitoring, fixes and operation are part of the monthly fee, so a vendor retirement is our problem to plan for rather than a surprise on your invoice. That is a feature of the service model, not a claim about results, and you should ask any provider, including us, to show you how they handled a retirement.
What does it cost to get a stalled project live?
Costs here are ranges built from stated assumptions, not quotes. Anyone who gives you a fixed number before seeing your systems is guessing.
The labor input. The U.S. Bureau of Labor Statistics reports that "the median annual wage for software developers was $135,980 in May 2025". Loading that at 1.3 to 1.4 times for payroll taxes, benefits and overhead gives roughly $177,000 to $190,000 a year, or about $85 to $92 an hour at 2,080 hours. Use your own blended rate if you have one.
The effort. The table covers the hardening work that comes after you already have a working pilot, if you do it yourself. It assumes the pilot's core logic is sound and the target systems have usable connections.
| Workstream | Typical effort | What drives the range |
|---|---|---|
| Test set and checking harness | 2-5 person-weeks | How much labeling real jobs requires, and whether someone who does the work has time |
| Data access, permissions and cleanup | 3-10 person-weeks | Number of source systems, and how much of the data lives in someone's spreadsheet |
| Access, safe writes, undo and audit log | 2-6 person-weeks | How many things it writes, and whether the target systems support reversal |
| Monitoring and cost reporting | 1-3 person-weeks | Whether you already have any |
| Security review | 1-4 person-weeks | Internal review only, versus a third-party test |
| Training and rollout | 1-4 person-weeks | Number of users, and how much the workflow changes for them |
At the low end that is 9 person-weeks of work, and at the high end 32. At $85 to $92 an hour and 40 hours a week, that puts a do-it-yourself production push in a band of roughly $31,000 to $118,000 in labor for a single workflow, before run costs. Multi-system projects land above it.
The run costs. Budget three lines that pilots never have: usage charges, which scale with volume; the human who handles what the automation cannot, a real salaried cost and often the largest of the three; and maintenance, which should assume one model migration a year plus ongoing checking.
The operated alternative. Fleurant AI's published prices are $18,000 to build three modules (scheduling and dispatch, invoicing and collections, and customer records, for example) plus a monthly fee set by company size: $2,000 up to 30 employees, $2,750 for 31 to 60 and $3,500 for 61 to 100. The band is reviewed once a year, never mid-term, so a new hire does not change the invoice. The first module goes live in 30 days and the full stack in 90, with each module running alongside the tool it replaces for two weeks before that subscription is cancelled. Whether that beats the do-it-yourself range depends on your software bill: below roughly $1,500 a month in subscriptions the arithmetic does not work, and the Stack Audit exists to show you your own numbers for $2,500, credited in full if you build within 90 days.
Key takeaway: Getting live commonly costs more than the pilot did, because the pilot bought a capability and going live buys trust, safety and the ability to keep it running. Teams that budget for one phase get stuck halfway through the second.
A 60-day plan to unstick a stalled project
If you have a demo that impressed people three months ago and has not moved since, this sequence tends to break the logjam. It assumes one workflow, one owner and a slow month.
Days 1 to 10: decide. Name the business owner and give them hours. Write the success numbers as numbers with a signature next to them. Write the renewal date and notice period for every subscription the project touches, and the date the old one gets cancelled. Define the failure level that triggers going back to the old way. Choose the autonomy level you will launch at, usually one level below where you expect to end up. None of this is technical, and all of it blocks everything else.
Days 11 to 25: prove. Build the test set from real jobs, checked by people who do the work, including the ugly ones. Publish the result even if it is bad, especially if it is bad, because a measured gap is a plan and an unmeasured one is an argument. If the numbers miss the threshold, shorten the chain before you reach for anything bigger. For how many checked cases you need before a score means anything, see our guide on how to evaluate AI chatbot accuracy.
Days 26 to 45: control. Give the system its own logins with the least access that works. Make every write logged and reversible. Stand up cost and run reporting. Build and test the kill switch. Put one or two technicians on the field side early, with real jobs.
Days 46 to 60: ship small and watch. Launch to one crew, one branch or one service line, with the exception queue staffed and a named person watching daily. Compare live numbers to the thresholds weekly, and run old and new side by side for two weeks before cancelling anything. Expand when the live numbers hold, not when the calendar says so.
Two notes on sequencing. Do not run these phases in parallel to save time, because phase 2 regularly changes the design that phase 3 would have hardened. And if phase 1 takes more than ten days, the problem is not the technology, and no amount of engineering will fix it.
When not to take a pilot live
Honest advice includes the cases where the answer is no.
When the workflow is rare. If the task happens fifty times a month and takes a person ten minutes, the automation saves about eight hours a month and costs more than that to operate and govern. Automate the thing that happens ten thousand times.
When nobody will own it. Software without an owner degrades silently. If you cannot name the person, do not launch; the question is organizational, and shipping will not answer it.
When the measured reliability does not support any safe autonomy level. If it is not good enough to act and not trusted enough to advise, going live will only move the failure somewhere more expensive.
When the real problem is the process. A surprising share of stalled pilots are automating a workflow that should be deleted or simplified, and sometimes the right answer is not automation at all but better software, a trade-off we wrote about in when to replace spreadsheets with custom software.
When it is your peak month. If the only available launch date is the week the phones are ringing off the hook, wait. A delayed launch costs a few weeks. A broken dispatch board costs customers.
Next steps
Getting from AI pilot to production is a management problem wearing a technical costume. The projects that ship choose the use case on business value, measure results with more than one number, give the thing an owner, and plan the cutover around the calendar of the business. The ones that stall usually skipped a decision, not a sprint.
If you have a stalled project, do three things this week. Score it honestly against the ten gates above. Run the step-reliability math on your actual workflow length. And name the person who will own it, out loud and in writing.
If you would rather see your own numbers first, the Stack Audit lists every subscription you pay for, maps where jobs get retyped, and ends with a costed plan and a payback period; it is $2,500 and credited back if you build. If you already know what you want replaced, see how the Operations Stack works, or talk to a specialist about where your project is stuck.
If you'd like one clear, practical take on the software that runs your business each month, subscribe to the Fleurant AI Briefing: one email a month, unsubscribe any time.
Frequently asked questions
What percentage of AI pilots make it to production?
There is no single credible number, and you should be suspicious of articles that give you one. The best-documented figures come from S&P Global Market Intelligence, whose May 2025 survey of 1,006 IT and line-of-business professionals in North America and Europe found that the share of organizations abandoning the majority of their AI initiatives before production rose from 17% to 42% year over year, with an average of 46% of projects scrapped between proof of concept and broad adoption.
Why does a pilot work in the demo but fail in the real business?
A demo runs one clean job once, with a person steering it. The real business runs hundreds of messy jobs a week, unattended, in your busiest month. Three things change: errors compound across the steps of a workflow, real data is dirtier than the sample (customer names spelled four ways, notes typed with thumbs from a truck), and every action that writes something, such as sending an invoice, now needs permissions, a log, an undo and an owner.
How long does it take to take a pilot live at a contractor?
For one workflow with a clear owner and tools that have working connections, plan on 6 to 12 weeks after a working pilot. Projects that need data cleanup, a new field-app rollout or sign-off from several people routinely run 4 to 9 months. The variable that moves the schedule most is not the software: it is how long it takes to get access to the real data and one named person who will own the result.
What should I measure before letting automation act on its own?
Measure what a person who does the job would measure: the share of jobs it handled correctly with no fixes, the share it escalated, how long a human still spends reviewing, and the cost per completed job. Track them on real jobs, not on the test set you built it with. Google Cloud's agent evaluation guidance also separates the final answer from the path taken to reach it, which is a useful habit for any multi-step workflow.
Do I need a governance policy before going live?
You need proportional controls, not a uniform policy. Gartner predicts that by 2027, 40% of enterprises will demote or decommission autonomous AI agents because of governance gaps found only after a production incident, and warns that applying the same controls to every agent causes both over-restriction and under-restriction. For a contractor the practical version is simple: read-only first, a person approves anything that writes, and a named owner for each automation.
What happens when the AI model my pilot was built on is retired?
It stops working. Microsoft sets an Azure OpenAI GA model's retirement date 18 months out at launch and returns HTTP 410 Gone afterwards, with at least 60 days of notice and no extensions. Anthropic commits to at least 60 days' notice before retiring a publicly released model. Treat the model as a replaceable part: keep a test set you can re-run, and budget for a migration roughly once a year, or buy it as part of an operated service.
How do I prove a pilot is worth taking live?
Measure it against a baseline of the same work done today: volume, handle time, loaded labor cost and error rate. Then report the share of items it fully handled, the review time it still needed, and the realized value after run costs. In IBM's 2025 CEO study, only 25% of AI initiatives had delivered the expected ROI. Our guide on the payback arithmetic for replacing software gives the formulas and a worksheet.
Should we build it ourselves or have a vendor's engineers build it for us?
Either can work, but decide the exit before you start. Gartner predicts that by 2028, 70% of enterprises will abandon agentic AI built by vendor forward-deployed engineering, and notes that these engagements often fail structurally before they fail technically. Contract for data export, a handover plan and clear ownership on day one, and measure success by whether someone other than the vendor can tell you how it works.
Sources
- Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, Gartner
- Gartner Says Applying Uniform Governance Across AI Agents Will Lead to Enterprise AI Agent Failure, Gartner
- Gartner Predicts 70% of Enterprises Will Abandon Agentic AI Built by Vendor Forward-Deployed Engineering by 2028, Gartner
- Gartner Says AI Projects in I&O Stall Ahead of Meaningful ROI Returns, Gartner
- IBM Study: CEOs Double Down on AI While Navigating Enterprise Hurdles, IBM Newsroom
- AI ROI: The paradox of rising investment and elusive returns, Deloitte
- Generative AI experiences rapid adoption, but with mixed outcomes - Highlights from VotE: AI & Machine Learning, S&P Global Market Intelligence
- Foundry Models lifecycle and support policy, Microsoft Learn
- Model deprecations, Anthropic
- Evaluate Gen AI agents, Google Cloud
- Define your success criteria, Anthropic
- OWASP Top 10 for LLM Applications 2025, OWASP GenAI Security Project
- More Vulnerable than You Think: On the Stability of Tool-Integrated LLM Agents, arXiv
- Occupational Outlook Handbook: Software Developers, Quality Assurance Analysts, and Testers, U.S. Bureau of Labor Statistics