The vendor says six weeks. The consultant’s report says eighteen months. Meanwhile your team has been running first drafts through a chatbot since spring without asking anyone. If you own a firm of five to fifty people and want to plan AI adoption around real dates, the honest answer arrives in stages, and each stage has a recognisable time pattern.
The calendar is also where these projects fail. Evidence from enterprise research and owner-managed case work points the same way. Firms that pace adoption around fixed decision points reach production. Firms that let pilots run open-ended become part of a very large failure statistic. Here is what the stages look like.
What are the typical timelines for AI pilots, rollouts, and scaling?
A proof of concept typically takes days to a few weeks. A pilot, where a small group uses the tool on live work, commonly runs eight to sixteen weeks. Rolling a successful pilot out across the firm takes three to six months more. Genuinely scaled AI, embedded in core operations, tends to arrive twelve to twenty-four months after you start.
The boundaries between those stages matter more than the labels suggest. Agility at Scale draws the useful line. A proof of concept tests technical feasibility, whether the tool can do the thing at all on sample data. A pilot tests real-world readiness, with live work, real users and real oversight. Conflate the two and you end up expecting production results from a fortnight of tinkering. Helium 42’s implementation roadmap places the pilot phase at eight to sixteen weeks inside an overall arc of six to eighteen months.
The upper end of the range is real. A peer-reviewed study of AI implementation in manufacturing firms tracked adoption across 21 months, from first exploration through prototypes to gradual scaling across machines and sites. Where AI touches physical processes or safety, one to two years is normal. A services firm moves faster because there is less to integrate. A realistic arc for a first serious use case, from planning to “this is simply how we work now”, is four to nine months.
Why do these timelines matter for your business?
The timeline is where AI projects live or die. IDC research reported by CIO found that 88 per cent of AI proofs of concept never reached production, with only four in every 33 making the cut. The typical failure mode was a pilot drifting for months without a decision, consuming senior time that in an owner-managed firm is directly billable to clients.
The pattern repeats wherever anyone measures it. A March 2026 survey by Maven AGI of 650 technology leaders found 78 per cent had at least one AI agent pilot running, while only 14 per cent had scaled one into organisation-wide use. BCG’s 2024 global survey found 74 per cent of companies struggling to achieve and scale value from AI despite widespread experimentation.
For an owner-managed business the stakes are sharper than the enterprise studies imply. A pilot that drifts from six weeks to six months burns the diary of people who bill clients, and a visible failure leaves the team cynical about the next attempt. McKinsey’s State of AI research finds that value concentrates in organisations that operationalise, deploying, monitoring and iterating rather than experimenting endlessly. The practical translation for a firm this size is the decision gate. Fix a review date before the pilot starts, then decide go, adjust or stop, and hold to it.
Where will you actually meet these timelines?
You will meet these timelines in three places. In vendor pitches, where a quoted six-week build rarely includes the change management on your side. In your own software, where AI features can switch on in an afternoon. And in compliance work, where a first data protection impact assessment adds two to eight weeks the first time you do one.
Vendor pitches deserve the closest reading. A quoted build time usually covers design, integration and testing on the vendor’s side. It rarely covers yours, cleaning the data the tool will read, training the team, rewriting the procedures the tool now sits inside. The slow part of any rollout is behaviour change, and behaviour change runs on a scale of months rather than sprints.
Your existing software is the opposite case. AI features now appear inside accounting platforms, office suites and practice-management tools, and switching one on takes an afternoon. Informal adoption starts there long before any formal project does, which is why a sensible pilot often begins by putting guardrails around what staff are already doing.
Compliance is the third meeting point, and it is scheduled work rather than a blocker. The ICO’s guidance on AI and data protection expects a data protection impact assessment before higher-risk processing begins, along with updated privacy notices and a way to explain AI-assisted decisions to the people they affect. The NCSC asks you to treat an AI service like any other third-party supplier, with due diligence, access controls and logging, which adds a few weeks of design time when client-confidential data is involved. The UK’s pro-innovation approach to AI regulation means no AI-specific licence is needed, but the existing rules apply from day one.
When should you plan in months, and when can you move in weeks?
Plan in months when the work touches personal data at scale, joins several systems together, or sits in a regulated sector. Plan in weeks when the scope is one internal task, the tool is packaged rather than built, and no client data is involved. The same firm can be in both lanes at once, and usually should be.
The months lane has four reliable markers. You are processing personal data at scale, customer records or HR files, which brings the ICO’s impact-assessment expectations into play. You are joining systems together, CRM, case management, phone logs, and the data needs cleaning first. You are building or heavily customising rather than configuring. Or you operate in a regulated sector, where FCA expectations or the EU AI Act’s high-risk categories can add several months of documentation and testing. Hit any of these and four to nine months is honest planning, not pessimism.
The weeks lane is narrower but genuinely fast. Keep the scope to one internal task, drafting documents, summarising meetings, internal knowledge search. Use a packaged tool rather than a build. Keep client and personal data out of the early work, and run the whole thing as a labelled experiment with opt-in users and human review. Under those conditions, two to eight weeks from idea to live use is achievable.
One caution on whichever lane you choose. A feature can appear inside software you already pay for and make a planned build redundant, and staff resistance can stretch even the simplest rollout well past its finish line. Treat any timeline as a working assumption you revisit at each gate rather than a promise you defend.
Which related concepts are worth knowing?
Four terms will keep coming up around AI timelines. Pilot purgatory describes projects that neither ship nor die. A decision gate is a scheduled point where you commit, adjust or stop. A proof of concept tests whether something works at all, and a data protection impact assessment is the ICO’s structured check before higher-risk processing begins.
Pilot purgatory sits behind IDC’s numbers, too many pilots, low conversion, and nobody willing to call time. If a pilot in your firm has slipped past its review date twice, it is in purgatory, and the kindest thing you can do is decide.
Decision gates are the antidote. A workable cadence for a firm this size is a two-week proof of concept, a six-week pilot with named users and written success criteria, a compliance check in week four, and a go-or-stop review in week eight. If the pilot clears the gate, allow three to six months of rollout to train the wider team and fold the tool into standard procedures.
The Monday move is smaller than the vocabulary suggests. Pick one task that visibly eats staff hours every week. Choose one packaged tool that plausibly addresses it. Put the week-eight review in the diary before you start, because the date protects you from the 88 per cent more reliably than enthusiasm does. And if you would rather pressure-test the plan before committing a quarter of your attention to it, book a conversation.
