Software Costs

How Long Does It Take to Build an AI Agent? 30 Days to Live

How long does it take to build an AI agent for a service business? 30 days to the first module, 90 for the full stack, and the seven delays that slip it.

Two colleagues planning a project on a large office whiteboard covered in sticky notes and a weekly schedule
In this article
  1. Why published timelines are close to useless
  2. The three clocks: person-weeks, calendar weeks and decision weeks
  3. How long does it take to build an AI agent? Four scopes, four answers
  4. What a 90-day build looks like, module by module
  5. Seven things that make these projects slip
  6. Estimate your own timeline with a six-question check
  7. How to compress the timeline without breaking things
  8. What can go wrong in the software itself
  9. When the honest answer is "not yet"
  10. A readiness checklist before you sign
  11. Putting a date on the calendar
  12. Frequently asked questions
  13. Sources

Short answer: how long does it take to build an AI agent for a service business? A working prototype on your real data takes one to two weeks. One agent doing one job properly, such as raising the invoice when a technician closes a job, takes eight to sixteen weeks if you build it yourself. If the agent is part of replacing your software, Fleurant AI commits to the first module live in 30 days and the full stack in 90. The model you pick changes those numbers by days. Getting at your data and getting your people to approve things changes them by months.

Most timeline articles give a table and stop. This one is written for the owner or general manager of a contracting company with a dispatcher, an office team and a software bill that grows with every truck. It shows where the weeks actually go, so you can check any quote against your own situation. You will get:

  • the four common scopes and how long each takes;
  • a 90-day schedule for replacing the software that runs the business, one module at a time;
  • the seven things that make these projects slip, and what to do about each;
  • a six-question check that takes most of the risk out of the first week;
  • the cases where the honest answer is "not yet".

"AI agent" here means software that does a piece of office work with some judgment: reading a technician's notes and drafting the invoice, matching a photo to the right job, chasing an unpaid bill. Most of the schedule has nothing to do with the AI part. It is the same schedule as any change to the system your office runs on.

Why published timelines are close to useless

Every vendor publishes a number, and each number is the vendor's best case. There is no independent survey of how long these builds take, so the honest approach is to read the vendors' own figures as a range and read the warnings alongside them.

For switching field service software, which is the closest thing to a benchmark, the published figures are:

Source Published timeline What it says drives delay
ServiceTitan (consultant interview) 12 to 16 weeks, with about four weeks of data discovery and four weeks of support after launch Data discovery
ServiceTrade 8 to 12 weeks on average Undated data migration, no written milestones, one-time training
FieldRoutes 2 to 3 months, with data import and cleanup in weeks 3 to 8 Data cleanup, running two systems at once

These are vendors describing their own products, so treat them as directional. What matters is that the three agree on the shape: a month or two of getting data ready, a few weeks of go-live, and a month of stabilizing. ServiceTrade also warns that a slow rollout can turn "an 8-week implementation into a 6-month one", and names the cause: customer data pushed to "phase two", and phase two never gets a date.

Gartner makes the same point about agent projects in general. It predicts 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", and notes that integrating agents into legacy systems "can be technically complex, often disrupting workflows and requiring costly modifications". Your field service platform and your accounting package are exactly those systems.

Key takeaway: If a vendor quotes a delivery date before asking who owns your data exports and who can approve a change to dispatch, the date is decoration.

Testing that before you sign is a solvable problem. Our guide to how to hire someone to build and run your software has the screening questions.

The three clocks: person-weeks, calendar weeks and decision weeks

Estimates go wrong when three different measurements get mixed up.

Person-weeks are engineering effort, and they are what a quote is priced on. The Bureau of Labor Statistics reports a median annual wage of $135,980 for software developers as of May 2025. The money side is in our guide to what it costs to build, run and maintain office software; this article is about the clock.

Calendar weeks are elapsed time. Two engineers do not halve a 12-week schedule, because much of the schedule is not engineering.

Decision weeks are the stretches when the build is waiting on one of your people: a login nobody can find, a question about how you price a second visit, a review the office manager cannot attend because it is the first hot week of the summer. Those are invisible in every estimate, and they are usually the difference between a project that lands in three months and one that lands in seven.

For a first build at a company that has never done one, engineering is roughly half of elapsed time. The rest is waiting and deciding. By the second or third module the access, the data and the habit of reviewing already exist, which is why the first module is slow and the third one is not.

How long does it take to build an AI agent? Four scopes, four answers

Agents are not one thing. The table gives four scopes, with engineering effort and elapsed time assuming one or two engineers working steadily and a responsive office. These are Fleurant AI planning estimates, not industry statistics.

Scope What it does at a contractor Person-weeks Calendar time Main constraint
Prototype Drafts invoices from job notes on 20 to 50 real past jobs, read-only 2 to 4 1 to 2 weeks Getting a data extract
Read-only helper Searches job history and drafts answers; a person sends everything 6 to 10 6 to 10 weeks Cleaning the records
One workflow that acts Creates the invoice, files the photos or sends the reminder itself 12 to 24 8 to 16 weeks Write access, approvals, a log of every action
A connected or replaced stack Dispatch, quoting, invoicing, records and the field app as one system Module by module 30 days to the first module, 90 to the stack Data export, cutover, the technician app

Three things stand out.

The prototype is fast, and that is the point. A prototype that works on real jobs in two weeks tells you whether the work can be automated at all, before anyone commits a budget. It is the cheapest way to find out you were wrong.

The jump from "reads" to "acts" is the expensive one. The moment the software can change something, you inherit permissions, approvals, an audit trail and a way to undo it. Many projects discover this in week nine of a ten-week plan.

Calendar time and effort drift apart as scope grows. At prototype size they are nearly identical. At production size, elapsed time is roughly one and a half to two times effort.

If you only need your existing tools to share data, a Connector is a smaller project than a replacement: $7,500 to build and $750 a month, live in two to four weeks. If you want to replace subscriptions, that is the Operations Stack.

What a 90-day build looks like, module by module

The reason the Operations Stack is built one module at a time is the clock. A module is a slice you can run, check and cut over by itself, so a problem in quoting never takes down dispatch. The commitment is the first module live in 30 days and the full stack in 90. The schedule below is how those weeks are typically used.

Schedule chart showing a 90-day build: data access days 0 to 20, module 1 live by day 30, modules 2 and 3 overlapping, the new system running beside the old one, and cutover by day 90
An illustrative three-module schedule. The blue bar is time you control, which is why it starts on day zero.

Before day zero: the Stack Audit

The Stack Audit takes one to two weeks and costs $2,500, credited in full against a build started within 90 days. It produces the list of every subscription, the workflow map of how a job travels from call to cash, and the overlap list. In schedule terms it is the cheapest way to remove the unknowns that cause the first slip.

Days 0 to 20: access and exports

This is the part of the schedule that belongs to you. The build needs an admin login for every system involved and an export of customers, jobs, price book, equipment records and service agreements. Request all of them on day one. Do not wait until the code works.

Days 5 to 30: module one, live

The first module is usually scheduling and dispatch, because it is where the per-seat bill and the daily friction are. It goes live on a slice of the work first. The schedule is fixed at 30 days; what flexes is the size of the slice.

Days 25 to 75: modules two and three

Quoting and invoicing, then customer records and the field app. They overlap because the access, the data and the reviewing habit already exist by then. Each is cut over on its own planned day, with a way back.

Days 30 to 80: run beside the old system

The new module and the old tool overlap on purpose. FieldRoutes advises against running two systems for the same work cycle, and the sensible reading is to split the work, not duplicate it: one trade or one branch on the new module, the rest on the old, and never the same job in both.

Days 75 to 90: cutover, then cancel

The last step is the one owners forget. The old subscriptions only stop costing money when someone cancels them, and contract end dates matter. Nobody cancelling the old tool is a delay in itself, and it is on the list below.

Key takeaway: Ninety days is realistic when the work is split into modules, the exports are ready in the first three weeks and one person on your side can decide. It is not realistic if any of those three is missing.

Seven things that make these projects slip

Almost every overrun traces back to one of these. Each has a fix that costs nothing if you apply it early.

1. The data is not ready. Duplicate customers, five spellings of one property address, a price book with items nobody has sold since 2019. ServiceTitan's own consultant describes about four weeks of data discovery for an implementation, and FieldRoutes puts import and cleanup in weeks 3 to 8. Fix: start the export on day one and agree who decides what a duplicate is.

2. Data migration gets postponed. ServiceTrade's warning bears repeating: data pushed to "phase two" never gets a date. Fix: put the migration on the schedule with a date, before kickoff.

3. Nobody on your side can say yes. One named person with authority, who answers within a day, saves more time than any tool. A committee of three owners who meet monthly adds a month to every decision.

4. It starts in the wrong month. A build that begins the week the season starts has no one to review it. Fix: start in the slow season, or finish module one before the peak.

5. Training happens once. ServiceTrade notes that technicians "forget half" of one-time training by the time they run live jobs. Fix: short sessions on the real field app on real jobs, and a person they can text.

6. You built an agent where a fixed workflow would do. Anthropic's engineering team advises finding the simplest solution possible, and increasing complexity only when needed. Chasing an unpaid invoice on a schedule needs no judgment at all. Fix: ship the fixed workflow first and add a model only where it cannot cope.

7. Nobody cancels the old subscription. The replacement is live and the old tool is still billing. Fix: put each contract's end date on the schedule, and the notice period next to it.

Estimate your own timeline with a six-question check

Borrowed averages are worthless. Answer these for your business, and count the answers that are "no" or "not sure".

  1. Can you export customers, jobs and the price book from your current system today, without asking the vendor?
  2. Do you have an admin login for every system involved?
  3. Is there one named person who can approve a change to dispatch, and will they answer within a day?
  4. Does your start date fall outside the six weeks before your busiest season?
  5. Can two of your dispatchers describe how a job moves from call to cash in the same way?
  6. Do you know the cancellation date and notice period of each tool you plan to replace?

0 or 1 "no": the 30-day and 90-day commitments are realistic. 2 or 3: expect module one to take the full 30 days and the stack to use all 90, and fix the weakest answers first. 4 or more: this is a real project, but spend the first two weeks on the list, not on building. A Stack Audit is built for exactly this: it ends with those six answers on paper.

How to compress the timeline without breaking things

Five things genuinely shorten a build, and some popular advice does not.

Start the waiting on day one. Logins, exports and cancellation dates have lead times measured in weeks. Requested on day one, they resolve while the build proceeds. Requested in week nine, they become the critical path.

Ship a thin slice, not a broad one. One module on one trade, end to end, beats three modules at 80%. A live slice produces the feedback that makes everything after it faster.

Stage the autonomy. Read, then draft for a person, then act with approval, then act within limits. Each stage is a shippable product. If the project has stalled after a promising demo, our ten-gate scorecard for getting a pilot into production covers what is usually missing.

Keep the rollback. Every cutover has a defined way back, tested once. It makes a bad day an inconvenience instead of a crisis.

Cut over by module, not by company. One Monday for the whole company is the pattern that blows up dispatch in August.

What does not work: adding engineers to a project that is waiting on your approvals, skipping the pilot slice to save two weeks, and choosing a more capable model to compensate for a workflow nobody has written down.

What can go wrong in the software itself

An agent that can change things needs limits. OWASP's LLM01:2025 Prompt Injection entry says it is unclear whether fool-proof prevention exists, and its mitigations are design decisions: give the software only the access it needs, require a person to approve high-risk actions, and keep untrusted content, such as a customer's email, away from the instructions.

For a contractor this is concrete. The software that raises invoices should not be able to delete a customer, and a customer message should not be able to change a price. Decide those limits in week one, because discovering them in week eleven means a rebuild. The practical rule is that anything touching money or the schedule gets an approval step until it has earned its freedom.

When the honest answer is "not yet"

Some projects should not start, and saying so early is worth more than any schedule.

You spend less than about $1,500 a month on software. Below that, replacing it does not pay back in a sensible time. A shop at that size is better served by tidying what it has. Our guide to what software can take off a small office team covers the cheaper routes.

The process is genuinely undefined. If two experienced dispatchers handle the same call differently and both are right, software will not settle it. Write the rules down first; that is a business project, not a software one.

The records live in spreadsheets and people's heads. That is a data project with software at the end. The test is in our guide to replacing spreadsheets with custom software.

You cannot name the decision maker. The project will stall at week ten however well it works.

You need the answer this quarter. If the honest estimate is 90 days and you need results in six weeks, a Connector or a prototype that informs the decision is a better use of the time.

A readiness checklist before you sign

Use this as the gate before anything goes live. A module should not cut over until every line has an owner and an answer.

  • Accuracy. A documented result on a few hundred real past jobs, with the sample size stated, at a bar agreed before the test.
  • Failure behavior. When the software is unsure, it asks a person. Verified, not assumed.
  • Permissions. The least access that works, per role.
  • Approval. A person approves anything that moves money or changes the schedule.
  • Audit trail. Every action logged: what it did, to which record, when.
  • Rollback. A tested way to switch it off and reverse what it changed.
  • Ownership. A named person on your side and a review date.
  • Exit. Your data exports in standard formats at any time, and the contract says what happens when you leave.

If a vendor's plan does not include these, the plan is shorter than the project. Support is the part owners forget to ask about. Fleurant AI's commitment is next business day for normal issues and four hours for urgent ones, with no overnight coverage, and it is better to know that before the first emergency than during it.

Putting a date on the calendar

If you need one planning number: for one workflow at a company that has never done this before, plan about twelve weeks from kickoff to production, and expect the critical path to run through your data and your approvals rather than the code. For a module-by-module replacement, plan the first module live at day 30 and the full stack at day 90.

Then do three things this week. List every system you pay for, with the cancellation date of each. Name the person who can approve a change. And request the data export, because that is the item with the longest lead time on the schedule.

If you want the list and the workflow map done for you, with the arithmetic on your own numbers, that is what the Stack Audit is: $2,500, one to two weeks, credited if you build. Or talk to a specialist first. A short conversation usually shows which of the six questions is the problem.

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

How long does it take to build an AI agent?

A working prototype on your real data takes one to two weeks. One agent doing one job in production, such as raising invoices from closed jobs, takes eight to sixteen weeks when you build it yourself. If you buy it as part of a replacement for your existing software, Fleurant AI's commitment is the first module live in 30 days and the full stack in 90. The model barely changes either number: data access and your own approvals do.

How long does it take to switch field service software?

Vendors publish 8 to 16 weeks for their own platforms. ServiceTitan's consultant interview says its implementation lasts 12 to 16 weeks, ServiceTrade says its contractors go live in 8 to 12 weeks on average, and FieldRoutes says a migration takes 2 to 3 months. Those are vendor figures, not independent benchmarks. All of them put data cleanup at the centre of the schedule.

Can the new system run beside the old one?

Yes, and for dispatch it should. Run the new module on a slice of the work, such as one trade or one branch, while the old system stays live for the rest. The one thing to avoid is two systems for the same job on the same day, which is how technicians end up not knowing where to look. Cut over one module at a time, not the whole company on one Monday.

What is the worst time of year to start?

The week your busiest season starts. Spring for landscaping and roofing, the first hot week for HVAC and the first freeze for heating and plumbing all leave the office with no spare hours to answer questions. Start the build in your slow season, or at least finish module one before the peak arrives. A build that stalls because nobody can attend a 30-minute review is the most common delay.

What should we have ready before day one?

A list of every system the business pays for, an admin login for each one, and a way to export your customers, jobs, price book and service agreements. Add one named person who can say yes to a change. Those three things decide the schedule more than any technical choice. The $2,500 Stack Audit produces the list and the workflow map for you.

Do we have to stop work during the build?

No. Dispatch, invoicing and collections keep running on your current tools while the new module is built and tested alongside them. Accounting, payroll and payment processing are never replaced, so those never move. What changes is who types what, and the cutover for each module is a planned day with a rollback, not an overnight switch.

Sources

  1. Mastering ServiceTitan: interview on implementation, ServiceTitan
  2. Why field service software implementations stall before they ever go live, ServiceTrade
  3. FAQ: Making the transition to FieldRoutes as your software provider, FieldRoutes
  4. Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, Gartner
  5. Building effective agents, Anthropic
  6. LLM01:2025 Prompt Injection, OWASP Gen AI Security Project
  7. Software Developers, Quality Assurance Analysts, and Testers: Occupational Outlook Handbook, U.S. Bureau of Labor Statistics

Free, no-obligation consultation

Have a question about software build timelines?

Tell us what you're working on or what you'd like to know. A specialist will get back to you with practical next steps, whether or not we end up working together.

  1. 1Send your question or project details (takes 2 minutes)
  2. 2A specialist reviews it and replies within 1 business day
  3. 3Get clear, practical next steps, free