AI Agent Development: Stages, Timelines, and What We Need From You

What you actually buy from an AI agent development company

Two things get sold under the same name, and they are priced the same way.

A platform brings its own scenario, and your process has to fit inside it. A custom build runs the other way around: the process you already have becomes the specification, exceptions included.

Neither is the wrong answer. If your process is standard, a platform will cover it faster and for less money. The trouble starts when one is sold as the other.

Gartner counted the market in June 2025 and found roughly 130 genuine agentic vendors among the thousands claiming the label. The rest relabelled assistants, chatbots and robotic process automation. A company buys “an agent” and receives a scripted bot with a new name and an annual subscription.

Three questions separate them during the first conversation.

• What happens to the exception? Ask about the case that comes up once a week and every employee knows by heart. “We will configure a rule” means a platform. “Show us twenty of those” means a build.
• Who owns the knowledge base? If the answers live inside someone else’s service and cannot be exported, you are renting a dependency.
• What happens when the process changes? A platform answers “you reconfigure it yourself”, which holds until the first change nobody planned for.

If those answers push you toward a simple bot, that is a fair outcome and a cheap one to reach before the work starts.

The four stages and how long each one runs

One note before the numbers. These come from our own projects. We looked for external measurements of agent delivery timelines and did not find any that name a source: search results give either vendor promises or figures repeated without an origin.

The order matters more than the durations. A small working version on one process comes first, everything else second. An ai agent implementation that starts with the full scope usually stalls on the same question: nobody can say whether things improved.

Notice where the decision point sits. The pilot, rather than the review call, is what answers “do we continue”. Until an agent has run on your data, an honest answer to that question does not exist.

So what pushes a project past these numbers? Almost never the technology. It slips when the requirements were described loosely and the wish list grows at every meeting. The fix is a person with the authority to say “that is not in the pilot, we will revisit it in version two”.

Stage one: the review call, where we decide whether you need an agent

The first call runs 15 to 30 minutes and ends one of two ways: a task for the pilot, or an honest “an agent will not help here”.

The second outcome sounds strange coming from a vendor, and it happens regularly. If a rule fits in one sentence, ordinary automation does the job for less money and with fewer things to check afterwards. An agent earns its place where the rule cannot be written down but examples of correct decisions exist: sorting free text, reading documents that arrive in four formats, matching records a person matches by eye.

Around half of what people bring us is faster to close with ordinary automation. We say so before the invoice.

We ask about three things on that call: what the process is, how many times a day it repeats, and how an employee knows they got it right. The third question is the awkward one and the useful one, because the answer decides whether the result can be measured at all.

Nothing is required from you at this point. No data, no access, no preparation.

Stage two: the pilot, and the part that depends on you

A pilot is a small working version on one process and on your real data. That is where the project first depends on you rather than on us.

Four things, and it helps to know about them early.

• An export of the process as a file. Several months of it, in a form somebody is allowed to send. A dashboard nobody can export from does not count.
• Read access to the system holding the data. Not a screen share. Permission to query.
• Labelled examples. A classifier usually needs a few thousand examples with a few hundred labelled by hand. Labelling a hundred takes two or three days of dull work.
• A person who accepts the result. Not whoever holds the budget. The one who looks at an answer and says it is correct, and still says so a quarter later.

Labelling is the hardest of the four, and here is why. Ask three employees to sort the same batch of requests and you get three different results, each defended confidently. Until those disagreements are said out loud and written down, the agent has nothing to check itself against: it learns whichever version of the truth appeared most often in the examples.

Your employee does the labelling while we set up the format. The correct answer lives with the person who runs the process daily, and examples are the only way to hand that knowledge over.

Do not let the list worry you. Nobody has perfect data, and what has to be ready before any AI project comes down to a shorter list than most people expect. Harvard Business Review and Cloudera surveyed more than two hundred executives in autumn 2025, and seven percent called their data fully ready for AI. Around a quarter said it was not ready or barely ready, and close to three quarters described preparing it as hard work.

A pilot ends with a working loop on one process and a number you can compare against the one before it. If that earlier number never existed, counting it is the first week of work rather than lost time.

Stage three: the build, where schedules actually slip

Timelines slip on access rather than on the model. The third stage shows it most clearly.

An export is promised in a week and arrives in six. The delay is worth breaking down, because almost none of it is technical.

First it turns out who really owns the data, and it is rarely the department that asked for the agent. Then the export is done by hand by one specific person with their own queue. Then security needs to understand what leaves the perimeter and under what terms. Each step takes days, together they take weeks, and nobody is at fault at any of them.

So we ask for access on day one. Put six weeks in your plan instead of one, and if it arrives sooner the project runs ahead of schedule.

It is worth knowing what stops projects like these. Gartner forecasts that over forty percent of agentic AI projects will be cancelled by the end of 2027 and names three causes: rising costs, unclear value, weak risk control. None of them is technical, which matches what we see in pilots that never reach production.

On our side the pilot grows three things. Integrations, so the agent reads and writes in your systems instead of living in a separate window. Exception handling, for the cases that come up weekly and went unmentioned during the pilot. Permissions, covering what the agent may do on its own and what it may only propose.

That last one deserves a sentence. Sending a client an email on the company’s behalf and drafting the same email are different levels of trust, and we separate them from the start. The choice follows the cost of a mistake rather than what is technically possible.

Stage four: support, or the agent quietly goes stale

You will not need a dedicated employee for support, and you will not be able to forget about the agent either. How long before things drift is impossible to state honestly: some processes hold for half a year, others gain a new form field a month after launch.

What the drift looks like is predictable. Quality does not collapse. It slides: the agent keeps sorting requests by rules that were correct in June, and nobody is watching.

An agent does not break loudly. It answers by yesterday’s rules, and there is nobody to notice.

The gap between “launched” and “still running” shows up in the large surveys too. McKinsey reports that close to two thirds of organisations experiment with agents while about a quarter scale one somewhere, and inside any single business function the share scaling stays under a tenth. Plenty explains a gap that size, and unbudgeted support is part of it.

A very ordinary example. Sales adds a “source” field to the request form, and the agent knows nothing about it. Reports still generate, everything looks fine, and it surfaces a month later when a manager cannot work out why one channel stopped producing leads.

Monthly work covers the quality of answers, retraining on new data, and whatever the process grew in the meantime. One person on your side who reads the report and says what changed is enough. The report shows which requests the agent gets wrong most often and which new types appeared.

This is why ai agent development services end with somebody continuing to answer for the system. We built that into how we work on purpose: a vendor who ships the code and leaves hands the client a problem instead of a result.

How to compare two vendors on the same task

By this point the questions worth asking are specific enough to compare answers side by side.

• What does your pilot end with? A working loop and a number, or a presentation.
• What do you need from us, and when? A vendor who has done this names the export, the access and the person who signs off, and names them in week one.
• What happens after launch? An ai agent implementation roadmap that stops at the launch date is missing its longest section.
• When would you tell us not to do this? Everyone has cases where an agent is the wrong tool. A vendor who has never met one has not been looking.

Any ai agent development company can show you a demo. The differences appear in the answer to the second question, because that is where a schedule turns into a plan you can actually hold someone to.

Where to start

Before you talk to anyone, check two things. Can your process be exported as a file covering several months, and is somebody allowed to send it? Who in the company will say “this works correctly” and stand behind it a quarter from now?

With both answers in hand, expect two to three months. Without them, start there, and an ai agent implementation plan will move faster than it does for companies working this out mid-project.

If you want a specific process looked at, book a free review call: the 15 to 30 minutes described above, where we look at the process and the data and say where an agent pays off and where the task has a simpler answer.