How long does it take to put an AI agent into production at a small business?
At Saména, a first agent is running one month after the call. French vendors that publish their timelines quote 3 to 16 weeks for a first agent or a proof of concept, depending on the state of your data, the number of tools to connect and how available your team is to review the work (ia.agency, Tensoria, Hack'celeration and Adevweb, pages read on November 19, 2026). A demo prototype is not yet production, and the gap between the two is why so many pilots stall.
Published by Saména
What timelines do vendors quote?
Six pages, all read on November 19, 2026, publish these timelines. They are vendors’ own claims with no independent measurement behind them, so read them as orders of magnitude. The range runs from three weeks to eight months, mostly because each source means something different by “production”: one agent on one task, or a rollout across the whole company. Our one-month timeline is a claim of the same kind. What matters to you is the gain, which the trial then measures in euros.
| Source | Stated timeline | What it covers |
|---|---|---|
| ia.agency (page updated 08/20/2026) | 3 to 8 weeks | From kickoff to a working agent, depending on how complex the integrations are. |
| Tensoria (agents page, 2026) | 4 to 8 weeks | A first agent in production on a targeted scope, kickoff included. |
| Tensoria (article of 07/05/2026) | 8 to 16 weeks | A well-scoped full project, from the kickoff workshop to production. For an agent that uses several tools or connects to a CRM or an ERP, 12 to 16 weeks. |
| Hack'celeration (“AI agency” page, updated in 2026) | 3 to 8 weeks, then a few months | 3 to 8 weeks for a first serious prototype, a few months for a rollout at scale. |
| Adevweb (article of 04/16/2026) | 4 to 8 weeks | A proof of concept on a reduced scope, with nothing in production. |
| Salesforce, State of Agentic AI 2026 survey, relayed by Glean (page of 10/30/2026) | 8 months on average | Getting to production, according to 2,025 AI decision-makers answering for themselves; the page we read does not give the size of their companies. |
One month at Saména: what happens?
One month after the call, a first agent is running. Setup starts with the call, a free 45-minute video meeting. Access to your tools is granted in two clicks, with no password, read-only. With your leadership, we start from your accounts to pick the process that matters most, with a yearly target in euros and a first milestone.
The first agent prepares the work (a reply, a quote, a document) and a person at your company approves it before it goes out. The four-month trial is there to put it through its paces: a point person on your side, sessions with each team on their real files and a weekly check-in, to fix what doesn’t help until your teams use it every day.
What slows other vendors down applies to us too. We look at it on the call.
What makes the timeline vary?
According to Tensoria, the variable is almost never the technology. The factors it lists:
- Access to the data and its quality, including data locked inside a tool with no way to connect to it.
- How available the business people are to review intermediate results.
- Integration with the existing information system, often underestimated: according to Tensoria, it easily takes 30 to 50% of a project’s time.
- Security and GDPR review, which Tensoria recommends starting at kickoff.
- Scope: a single, narrow use case does the most to keep the schedule.
What are the phases of a project, and how long does each take?
These durations come from the sources we read; a phase with no published duration says so.
| Phase | What happens | Published duration |
|---|---|---|
| Kickoff | Choose the process, write down the cases covered and the cases excluded. | 1 to 2 weeks, according to Tensoria (07/05/2026). |
| Prototype or proof of concept | An agent on a reduced scope, nothing in production. | 3 to 8 weeks according to Hack'celeration, 4 to 8 weeks according to Adevweb; a first working pilot in 4 to 8 weeks according to Tensoria. |
| Pilot | Real people, your real documents, every output reviewed by someone from the business. | No published duration. According to Adevweb, manual review comes off only after 3 to 6 months of proven reliability. |
| Production | A restricted scope first, then widened, with an observation period. | It closes Tensoria’s 8 to 16 weeks; the observation period has no published duration. |
| Rollout at scale | The whole team, or several departments. | A few months, according to Hack'celeration. |
| Measuring the gain | Comparing before and after on the same processes. | 3 to 9 months for a scoped project according to Adevweb; 3 to 6 months for a single high-volume workflow according to Druid AI, relayed by Glean. |
Prototype, pilot, production: what is the difference?
The word “agent” covers three different things depending on where the project stands.
- A demo prototype shows that an idea works on a chosen case, often with a few clean documents. Nobody uses it every day. According to Adevweb, nothing is deployed in production at this stage.
- A pilot puts the agent in the hands of a few people, on your real documents, with their access rights. Someone from the business reviews every output. You measure reliability, time saved and how well the team takes to it.
- Production has the agent serving the whole team. Someone owns it, it is monitored and maintained (see what happens to an agent after delivery), and its gain is measured in euros.
Why do so many pilots stall?
Among French mid-sized companies, the barometer published by Bpifrance Le Lab (the research arm of the French public investment bank) on June 23, 2026 found that 77% of executives say their employees use generative AI at work, and that 17% of the companies using it see time savings (23% among regular users).
In July 2024, Gartner forecast that at least 30% of generative AI projects would be abandoned after the proof of concept by the end of 2025, because of poor data quality, inadequate risk controls, escalating costs or unclear business value. That is a forecast, not a count.
The MIT NANDA report “The GenAI Divide,” published in 2025 and relayed by Fortune, estimates that about 5% of enterprise AI pilots produce rapid revenue acceleration, with the rest showing no measurable effect on the bottom line. It blames flawed integration into the organization more than model quality. Tools bought from specialized vendors succeed about two thirds of the time, and internal builds a third as often. The study rests on 150 interviews, a survey of 350 employees and 300 public deployments, a modest base for such a sharp number. Saména sells this kind of project, so read that last result with that in mind.
A pilot built in three weeks with no review, no measurement and no owner will not reach production. The total timeline is decided in the move from prototype to production. To review a project that is still a prototype, see ten questions to review an AI spec.
What do you measure before committing?
During the four-month trial, the gain is measured in euros on your own processes, before any long-term commitment. The yearly target in euros and the first milestone are set at the start with your leadership, then tracked with them. After the trial, an all-inclusive subscription (hosting, maintenance, updates, support, uncapped usage), with unlimited users.
The sources we read put the payback of a single high-volume workflow between 3 and 6 months (Druid AI, relayed by Glean) and that of a scoped project between 3 and 9 months (Adevweb). So the trial covers the process that matters most, which fits in four months; according to Druid AI, a rollout across several departments takes 12 to 18 months. The method is laid out in How to measure the gain in euros, and public market prices are in How much does an AI agent cost a small business?.
How can you shorten the timeline on your side?
Four moves, drawn from the factors Tensoria cites:
- Pick one narrow task whose result you can put a number on.
- Name the person from the business who will review the output, and free up their time.
- Before the call, check where your documents are and who has access: a tool with no way to connect to it lengthens the project.
- Start the security and GDPR review at kickoff. Fourteen questions to ask an AI vendor before you sign give you a starting point.
Frequently asked
- How long does it take to put an AI agent into production at a small business?
- At Saména, a first agent is running one month after the call. French vendors that publish their timelines quote 3 to 16 weeks for a first agent or a proof of concept, depending on the integrations, according to ia.agency, Tensoria, Hack’celeration and Adevweb; a rollout at scale takes a few months.
- Can an AI agent be ready in two weeks?
- The sources we read do not go below three weeks for a serious prototype (Hack’celeration) or a working agent (ia.agency). A shorter promise deserves questions: on which documents, with what review, in production or as a demo?
- What is the difference between a prototype, a pilot and production?
- A prototype shows an idea works on a chosen case. A pilot puts the agent in the hands of a few people, on your real documents and with human review. In production, the agent serves the whole team, someone owns it and its gain is measured.
- Why do AI pilots not reach production?
- Gartner forecast that at least 30% of generative AI projects would be abandoned after the proof of concept by the end of 2025, in part because of poor data quality, escalating costs or unclear business value. Among French mid-sized companies, the Bpifrance Le Lab barometer of June 2026 found that 77% of executives say their employees use generative AI, and that 17% of the companies using it see time savings.
- What makes an AI agent project take longer?
- According to Tensoria: access to data and its quality, how available the business people are to review, integration with the information system (30 to 50% of a project’s time), security and GDPR review, and a scope that is too wide.
- How long before you see a gain?
- According to Adevweb, 3 to 9 months for a scoped project; according to Druid AI, relayed by Glean, 3 to 6 months for a single high-volume workflow. At Saména, a four-month trial measures the gain in euros before any long-term commitment.
- What happens during the first month at Saména?
- A call, access to your tools granted in two clicks and read-only, the choice with your leadership of the process that matters most, then a first agent running one month after the call. A person at your company approves before anything goes out.
Sources consulted
The pages we read to write this one, and the day we read each.
- ia.agency: AI agent agency (updated 08/20/2026, in French) (accessed )
- Tensoria: AI agents, offer and timeline (in French) (accessed )
- Tensoria: timelines of an AI project (article of 07/05/2026, in French) (accessed )
- Tensoria: time to return on investment of an AI project (article of 05/26/2026, in French) (accessed )
- Hack'celeration: AI agency (updated in 2026, in French) (accessed )
- Adevweb: AI automation for small businesses (article of 04/16/2026, in French) (accessed )
- Glean: how long before an AI agent pays off (updated 10/30/2026, relays Salesforce and Druid AI, in French) (accessed )
- Finyear: Bpifrance Le Lab ETI barometer 2026 (06/23/2026, in French) (accessed )
- Reuters, reprinted by The Star: French mid-sized firms adopt AI but see few gains (06/23/2026) (accessed )
- APMdigest: Gartner, 30% of generative AI projects abandoned after proof of concept (press release of 07/29/2024) (accessed )
- Fortune: MIT NANDA report on generative AI pilots (08/18/2025) (accessed )
