"You’ll have AI running in two weeks." It’s the line repeated most on sales calls and the one that buries the most projects. Technically it isn’t a lie —some things do go live in two weeks— and strategically it’s a trap: they sell you opening day, and you hear that in two weeks the problem is solved. Not the same thing. Let’s put the real timelines on the table, with a source, and separate what switches on fast from what takes a full season to deliver actual results.
How long it takes to implement AI in a company: it depends on what you call "AI"
There’s no single timeline for "implementing AI" because there’s no single "AI". There are two worlds with different clocks. An off-the-shelf automation —a chatbot on APIs that already exist, a flow that connects two apps, data extraction with a market model— reaches production in weeks. A custom system —a model tuned to your data, a RAG over your documentation, an agent with judgment and guardrails— takes 3 to 6 months to actually be in production. Confusing the two worlds is where almost all the disappointment comes from.
| Type of project | Real time to production | What controls the clock |
|---|---|---|
| Off-the-shelf SaaS tool | Days to 2 weeks | Configure and connect, little more |
| Off-the-shelf automation (chatbot, flow, extraction) | 6–10 weeks | Integration with your systems |
| Custom system (own model, RAG, agent with judgment) | 3–6 months | Data, integrations and edge cases |
| Company-wide AI rollout | 12–18 months | Governance, compliance and adoption |
I’m not making these ranges up. Aggregated market analysis puts the average enterprise AI rollout at 6 to 12 months from kickoff to production, with focused automations on pre-built platforms at 6 to 10 weeks and custom or company-wide LLM deployments at 12 to 18 months (Alice Labs, May 2026, citing McKinsey and Gartner). Translation: if someone promises you "AI" in two weeks, they’re either selling you the easy piece or not telling you the rest.
Why "AI in two weeks" sells launch day, not the season
The two-week promise is true for one very specific thing: switching on a demo. A chatbot on a market API stands up in days, and on screen it looks done. But switching on isn’t solving. Between the demo that works on Tuesday and the system that holds up in your real operation there’s a whole season: the data you have to clean up, the integrations with your old ERP, the edge cases that don’t show up until month three, and the people who have to change how they work. The demo is planting day. Results are the harvest, and the harvest doesn’t run on your sales calendar.
That gap between what switches on and what pays off is a sibling of the gap between what gets quoted and what ends up billed, which we take apart in how much it costs to implement AI in a company. Timeline and price hide behind the same trick: show the easy part and charge for the hard part by surprise.
What really moves the timeline (the four clocks)
If you want to estimate your own case before believing any calendar, look at these four clocks. They’re the ones that actually run the show, in order of impact:
- The state of your data. It’s the biggest clock and the one nobody looks at: data preparation eats 40% to 60% of a project’s total time (Gartner, via Alice Labs). If your data lives tidy, you start; if it lives in PDFs, emails and spreadsheets nobody documented, there are weeks of work before you touch the AI.
- The integrations. Connecting a modern system by API is fast. Connecting a 2004 ERP with custom middleware adds weeks that are almost never in the initial plan.
- The edge cases that show up in month three. No pilot sees them, because a pilot works with clean examples. The badly formatted invoice, the order that doesn’t fit, the exception that needs judgment: that only surfaces at real volume, and each one reopens the testing cycle.
- The change in people. The number-one delay cause isn’t technical: 42% of organizations name change management and adoption as their biggest brake (Deloitte, via Alice Labs). A system nobody uses isn’t in production, it’s in a screenshot.
The first two clocks are plumbing: getting your data and systems in shape so the AI can lean on them, which is what we do in enterprise AI infrastructure. The last two are the craft of implementation: scoping the process, handling the edge cases and leaving it running. How that’s done well we develop in automating a process with agents and in how to build an AI agent that actually works —spoiler: the "switching it on" part is the short one—.
Off-the-shelf vs custom: the three-to-five-month gap
The decision that moves your timeline most is one: pre-built platform or custom model. Organizations using pre-built platforms deploy 58% faster than those building a custom model, and that difference works out to 3 to 5 months of calendar (Gartner, via Alice Labs). The practical read: if a well-configured market platform solves your pain, building a custom model doesn’t make you smarter, it makes you slower. And the other way around: if your case is genuinely your own, that custom model is what brings the edge —but then accept the season, not the week. If the pain is moving data between the apps you already use, it’s almost always operations automation on top of what you have —weeks, not months—.
Going it alone vs with help: why "alone" runs slower
Going it alone isn’t free in time. The SMB that goes solo trips on exactly the two big clocks: it gets stuck in data preparation because nobody on the team has done it before, and it underestimates adoption because it isn’t their craft. That’s why 63% of organizations overrun the timeline they estimated, almost always by underestimating what it took to get the data ready (KPMG, via Alice Labs). A team that has already built your same case skips discovery and shaves off weeks: not because it works faster, but because it already knows where the mines are. The honest math isn’t "free vs expensive", it’s "months stumbling vs a few weeks with a safety net".
How to ask for a timeline and not get sold smoke
The way to break the game is to ask for the timeline in parts. Three lines that change the conversation: "give me the timeline to the first version in production, not to the demo"; "how much of that timeline is preparing my data?"; "what happens when the cases the pilot never saw show up?". Whoever implements answers all three without sweating, because they live off the system working in your operation. Whoever does theatre repeats "two weeks" and changes the subject, because they live off you signing before you look at the real calendar.