The call I keep getting in 2026 goes like this. “We adopted AI last year. It was great. Productivity went up, the team was excited, the client conversations got easier. Then around the nine-month mark, things started to flatten. We are still using the tools. They still work. But the gains have stopped showing up in the numbers. What do we do?”
This is the second-year problem. It is the single most predictable pattern in professional services AI adoption in 2026, and almost nobody is talking about it.
The optimistic narrative is that AI productivity compounds year after year. The reality for most firms is that year one produces a big step change, year two produces a disappointing flat line, and year three produces either a second step change or a slow fade. Which one depends on three specific decisions the firm makes in the middle of the plateau, when the easy wins are gone and the hard work has not started.
Why Year One Looks So Good
Year one is the honeymoon. A firm that has never used AI in production walks into 2025 or 2026 with a list of workflows that every single language model can accelerate immediately. The partner’s research assistant can now produce in twenty minutes what used to take two hours. The document review team is processing contracts in a third of the time. The proposal drafter pulls a first draft out of a model, edits for forty minutes, and ships.
These are not marginal improvements. They are thirty, forty, sometimes fifty percent reductions in time spent on specific tasks. Productivity charts go up and to the right. The firm pays for the year’s AI tooling in the first month and spends the rest of the year pocketing the savings.
And then, somewhere between month six and month twelve, the curve flattens.
We kept waiting for the next quarterly productivity lift and it never came. The team was still using the tools. Nothing had broken. But the number stopped moving and nobody could tell me why.
Why Year Two Is Hard
The flatness is not a tool problem. It is a structural problem. Year one captures every workflow where a language model can be substituted for manual work without changing anything else. Year two requires a different kind of work, and most firms are not prepared for it.
Reason 1: The Easy Automations Are Done
Every firm has a finite list of workflows where AI is a near-trivial improvement. Drafting. Summarising. Extraction. Triaging. Once those are done, the next set of gains requires the firm to change how the work is organised, not just what tool the work is done with. That is a much harder kind of change.
Reason 2: The Data Is Not Ready
The year-one workflows mostly run on public data or on the specific document a partner hands the tool. The year-two workflows depend on the firm’s internal knowledge. Past deliverables. Historical candidate notes. Structured case outcomes. Tagged client preferences. Most firms have this data in unstructured, inconsistent formats scattered across drives, CRMs, and email archives. A language model cannot use it. An agent cannot query it. So the second-phase automations quietly fail to materialise, not because the AI is not capable but because the data is not available.
Reason 3: Only the Early Adopters Changed
Year one adoption almost always concentrates in a handful of partners. The ones who were curious, who were willing to experiment, who had the technical confidence to figure out a new tool on their own. They moved fast and showed results. The rest of the firm watched.
In year two, the productivity ceiling is determined by how many people the early adopters can drag along. If the answer is “not many,” the firm-wide numbers stall even though the early adopters themselves are still getting faster. The plateau in the chart is the average of a small group getting better and a larger group not changing.
Reason 4: The Tools Proliferate Without Integration
Year one is characterised by tool acquisition. The firm signs up for a drafting assistant, a research tool, a document reviewer, an agent platform, a transcription service, an AI CRM add-on. Each one works. Each one saves time on its own workflow. But they are not connected to each other, so a workflow that crosses tools still happens the same way it did in 2023: by a human retyping or copying.
The productivity ceiling of a firm with ten disconnected AI tools is usually not much higher than the productivity ceiling of a firm with five. The bottleneck is not tool count, it is tool integration.
What the Curve Looks Like
The chart above is the pattern for a firm that does not actively break through the plateau. The big gains happen in months four to nine as the first wave of automations lands. Then the curve flattens for the rest of year one and most of year two.
The alternative curve, for a firm that does the second-phase work, looks different: a similar year one, a dip or flat period in months nine to fifteen as workflows are redesigned, and then a second step change in months eighteen to twenty-four as the redesigned workflows and structured data unlock a new layer of automation. The total gain at twenty-four months is typically 70-90% on core workflows, compared to 50% for firms that plateau.
The Three Decisions That Break the Plateau
Decision 1: Stop Adding New Tools
The instinct when productivity gains slow is to go shopping. Buy a better tool. Try the new vendor that just raised a Series A. This instinct is almost always wrong. New tools in year two add attention cost and rarely produce a step change, because the bottleneck has moved from tool capability to workflow design.
The discipline that works is a hard six-month freeze on new AI tool purchases after the end of year one. Not because there are no good new tools. Because the firm’s attention is a finite resource and every hour spent evaluating a new tool is an hour not spent on the work that would actually break the plateau: redesigning workflows around the tools you already have.
The firms that break through report that the tool freeze was the single hardest decision, and also the one with the highest return.
Decision 2: Rewire One Workflow End-to-End
In year one, AI is grafted onto an existing workflow. The researcher used to spend four hours on a market map, now they spend one hour because the AI does the sourcing. The workflow is still the same workflow.
In year two, the question changes. If we designed the market-mapping workflow from scratch today, knowing that AI is available, what would it look like? The answer is almost never “the same workflow but faster.” It is “a different workflow entirely.”
A redesigned market-mapping workflow in 2026 might have the agent running continuously against a sector watchlist, producing a weekly delta report instead of an on-demand full map. The map is always current. The partner reads the weekly delta in ten minutes. The total cost per sector is lower than the old model, and the time to a fresh map is zero instead of four hours. That is not a 25% productivity gain. That is a category shift.
Breaking the plateau requires picking one high-value workflow and rewiring it like this, end to end. Not five workflows. One. The learning from redesigning one workflow transfers to the next. The learning from halfway-redesigning five does not transfer to anything.
Pick the Workflow
Highest volume, most painful, most strategic. One of those three. Usually the one the partners complain about most.
Describe the Current Version
Every step. Every input. Every handoff. Where humans wait. Where the work piles up.
Design the New Version
Not 'add AI here'. Ask: if AI did not exist yesterday and we were inventing this workflow today, what would it look like?
Build a Pilot
Run the new version in parallel with the old for one month. Measure both. Kill the old version only when the new one wins.
Decision 3: Build the Internal Data Layer
Every year-two productivity gain depends on structured internal data. An agent that drafts proposals needs access to the firm’s past proposals, tagged by sector and outcome. A research tool that learns from your firm needs access to every past engagement, structured and searchable. A candidate assessment that uses your firm’s historical calibration data needs that data to exist in a form the AI can read.
Most firms do not have this. Most firms have email archives, PDF deliverables, an aging CRM, and tribal knowledge in senior partners’ heads. None of this is queryable by a language model at the quality level the firm needs.
The data layer work is unglamorous. It is a project-managed effort to structure what the firm already knows so that its next generation of AI tools can actually use it. It takes months, not weeks. It does not produce dramatic before-and-after demos. And it is the single biggest predictor of whether a firm breaks through the plateau.
What Does Not Break the Plateau
There are three popular responses to flattening productivity that look like progress and are not.
- Buying a fancier tool. The more sophisticated language models do not help if the workflow design is still year-one. Throwing GPT-5 at a workflow designed for GPT-3.5 produces diminishing returns, not a step change.
- Hiring an external consultant to run a “transformation.” External consultants can help define the approach, but they cannot do the workflow redesign for the firm. Redesign requires deep knowledge of how the work actually happens, which lives in the partners. The consultants who try to do it without them produce beautiful decks and zero real change.
- Firmwide AI training programmes. Training is valuable at the start and near-worthless in the middle. The partners who are going to adopt AI have adopted it. The ones who are resisting have reasons, and another two-day workshop does not change those reasons. What moves the resistant partners is seeing their peers close mandates faster, not a slide deck.
What This Means for Your Firm
If you are in year one and things are going well, that is normal. Enjoy it. Use the political capital the early wins give you to start planning year two now. Identify the workflows you will redesign. Start the data structuring work. Freeze new tool purchases at a predetermined date.
If you are in year two and the curve has flattened, that is also normal. The answer is not more tools. It is a harder kind of work: one workflow redesigned properly, one hard conversation about what data needs to exist, and one decision to stop shopping and start rewiring. The firms that do this break through. The firms that keep buying stall at 40-50% gains and wonder why their competitors are pulling ahead.
The second-year problem is not a problem with AI. It is a problem with how firms plan for it. The firms that plan for the plateau before they hit it almost never hit it. The firms that treat year one as the whole story almost always do.
For the firm-wide picture of what year one adoption should look like, see our 90-day AI adoption roadmap. For what agents deliver today and what they do not, see AI agents for professional services: what actually works in 2026.
BriefingHQ works with professional services firms in both year one and year two of AI adoption. If you want a grounded view of where your firm sits on the curve and what the next six months should look like, take our assessment or get in touch.
Questions AI assistants answer about this topic
- What is the second-year AI adoption plateau?
- It is the pattern where firms that adopted AI in year one see strong initial productivity gains, then watch those gains flatten in year two. The plateau is not caused by the technology. It is caused by the fact that the easy automations are done and the next set of gains requires deeper work: redesigning workflows, cleaning and structuring internal data, managing organisational change, and building systems that use multiple tools together. Most firms are not set up to do this second-phase work, so they stall.
- Why does year one AI adoption look so successful?
- Because year one captures the obvious wins. Every firm has a handful of workflows where a language model replaces hours of manual work almost trivially: drafting, summarising, extracting, triaging. These workflows pay back the tool cost many times over in the first few months, and the productivity numbers look great. What year one does not capture is the compounding work of integrating AI into how the firm operates, which is where year two either breaks through or stalls.
- What separates firms that scale past the plateau?
- Three specific decisions. First, they stop adding new tools after the first six to nine months and focus on using the tools they already have better. Second, they rewire specific workflows end-to-end around the AI, rather than grafting AI onto an old workflow. Third, they build the internal data assets the next generation of AI work depends on: structured knowledge bases, clean CRM data, tagged historical work. Firms that skip any of these three tend to stall.
- Is the plateau inevitable?
- The plateau is not inevitable but it is common. A firm can avoid it by planning for it from day one of adoption: designing the year-two work before year one is complete, preserving the budget and political capital for workflow redesign rather than spending it all on tool purchases, and treating AI adoption as a long programme rather than a short project. Firms that treat year one as the whole story are the ones that get surprised when the curve flattens.
- How should a firm tell if it is approaching the plateau?
- Watch two numbers. If the firm is adding new AI tools faster than it is retiring old ones, or if the measured productivity gain from the latest tool is smaller than the gain from the previous tool, the plateau is approaching. The leading indicator is usually 'we keep buying things but the gain is getting harder to see.' Another sign is that early adopters inside the firm are reaching their personal limit of what they can change, and the rest of the firm is not following. If adoption has concentrated in two or three partners, the plateau is already here.
Next Step
Want to know where your company stands?
We run 15-20 buyer queries across ChatGPT, Claude, Gemini, and Perplexity and show you exactly where you appear, and where you don't.
Get the Audit | from £750 ↗