We Have Too Many AI Tools and No One Uses Them Properly. Where Do We Start?
Start from the bottleneck, not the inventory. The instinct when tools pile up is to audit them, rationalise them, and govern them, and that instinct produces a tidier version of the same problem. The discipline that works is what we call the Bottleneck Principle: the unit of AI value is the human bottleneck, not the use case. Anything can be solved now; the hard part is finding the problems worth solving. So ask each team for the one piece of work that costs it most pain each week, pick the sharpest, and solve it properly with whichever single tool fits. Let the working solutions reveal the stack, then retire what nothing depends on. Gartner found only 13 per cent of organisations think they have the right AI governance in place, and I would argue governance follows focus, not the other way round.
Sprawl is the default, and it is about to get much worse
If your organisation holds more AI tools than anyone can name, you are not behind. You are the median. Licences arrived through IT, through functions buying their own, through vendors bundling AI into products you already had, and through people simply signing up. Nobody decided to build a sprawl. Sprawl is what happens when acquisition has no anchor.
The next wave dwarfs this one. Gartner predicts that by 2028 an average global Fortune 500 enterprise will have over 150,000 AI agents in use, up from fewer than 15 in 2025, and found that only 13 per cent of organisations think they have the right AI agent governance in place (Gartner, April 2026). The same analysts name the mechanism that defeats top-down control: when people cannot work in the sanctioned tools, they go around the controls into shadow AI. Locking the front door does not work when the building has no walls.
So the question "which tools should we keep" is about to become unanswerable by cataloguing, because the catalogue will be growing faster than the audit. The organisations that stay coherent through this will not be the ones with the best inventory. They will be the ones that know which problems they are solving.
The pilot arithmetic
Spreading across many tools and pilots feels like prudent option-taking. The measured results say it is how value dies. Gartner's survey of 782 infrastructure and operations leaders found only 28 per cent of AI use cases fully succeed against ROI expectations, with 20 per cent failing outright, and the most-cited cause in the leaders' own account was expecting too much, too fast (Gartner, April 2026). Its analysts' remedy reads like a rebuke of the sprawl itself: manage AI as a portfolio, avoid duplication, concentrate resources where they matter.
The same concentration shows up wherever adoption is measured at scale. Stanford's AI Index reports organisational adoption at 88 per cent, generative AI in at least one function at 70 per cent of organisations, and yet agent deployment still in the single digits across nearly every business function (Stanford HAI, 2026). Broad, shallow, and stalled at the point where a tool would have to actually run something. And Grant Thornton's survey of 950 senior leaders found the organisations pulling ahead are scaling fewer pilots, with better measurement and clearer exit criteria, not more; those with fully integrated AI report revenue growth several times that of the perpetual piloters (Grant Thornton, 2026).
Read together, the numbers describe one pattern. Breadth is cheap and everyone has bought it. Depth is scarce and the returns live there.
The inventory is not the asset
Here is the move that unsticks the conversation: stop treating the tool list as the thing to be fixed. A tool inventory is a record of purchasing decisions. It tells you what the organisation was sold, in what order, and by whom. It tells you nothing about where value is available, because value does not live in tools. It lives in the specific points where your people's work jams, and those points exist entirely independently of what is on the licence list.
Nobody has a tools problem. They have a problems problem: too many tools, and no named problems to point them at.
This is why consolidation-first fails. Cut thirty tools to eight and you have a shorter list that nobody uses properly, because the thing that was missing, the anchored problem, is still missing. It is also why the sprawl produced so little conflict on the way in: a tool that is not attached to anything important is easy to acquire and easy to ignore. The moment a team is solving a problem it actually owns, the tool question answers itself in about a day, and it usually has a one-word answer.
The Bottleneck Principle
The discipline we install for this is called the Bottleneck Principle, and it is one sentence: the unit of AI value is the human bottleneck, not the AI use case. Anything can be solved now. The capability question, "can AI do this?", is almost always yes, which is exactly why it stopped being a useful question. The hard part, the part that deserves the senior hours, is finding the problems: the recurring points where a team's work actually jams, where hours drain every week, where a decision waits on work nobody has capacity to do.
Two supporting rules make it operational. The first is bottleneck economics: a bottleneck is a cost, solving it is a return, and the return is weighed net, after the effort of solving. That is the investor's lens, and it ranks problems the way a use-case scoring matrix never can, because it prices what keeping the problem costs. The second is owning the clock: concentrate AI where the team controls the timeline, because work gated on outsiders under-yields however well the machine performs.
Notice what this does to the original question. "We have too many tools, where do we start" turns out to be the same question as "which problem do we solve first", and that question has an actual answer, discoverable in a working session per team. Finding it is the exercise we run in the Applied Deep Dive Workshop, and the reason the workshop works on the team's own problem rather than a case study is the whole argument of this page: the problem is the scarce input, not the tool.
The sequence, one step at a time
- 01
Name the bottlenecks, team by team.
Ask each team leader one question: what is the piece of work that costs your team the most pain, every week, that you own end to end? Not an AI idea. A pain. One working session per team surfaces it, and the leader who cannot name one has told you something useful too.
- 02
Rank them on bottleneck economics.
Price what keeping each bottleneck costs, in hours, money or stalled decisions, and weigh the return net of the effort to solve. Keep the list short. Two or three anchored problems beat a portfolio of twenty, on all the evidence above.
- 03
Solve one properly, with one tool.
Bring the machine to the jam with the reasoning discipline on, which means the team works it through the 4-Lines on live work rather than receiving a demo. Whichever tool the solution actually needs is, by definition, a keeper. Most solutions need one.
- 04
Read the result honestly.
Define the reading before you start and take it at eight weeks: the voluntary return, the output quality, the decisions that moved. The full set of honest readings is in how to know whether AI adoption is actually working. Graduate or stop; no pilot without an exit criterion.
- 05
Only now, rationalise the stack.
After a few anchored solutions are running, the working stack has revealed itself. Retire what nothing depends on. This is the same audit you were tempted to run first, except now it has a criterion other than opinion, and the savings fund the next bottleneck.
If your rolled-out assistant is part of the sprawl, the companion essay on what to do when Copilot changed nothing runs the same logic for the single-tool case. The sequence is identical because the disease is: acquisition without an anchored problem, at every scale.
Frequently asked questions
We have too many AI tools and no one uses them properly, where do we start?
Start from the bottleneck, not the inventory. The Bottleneck Principle says the unit of AI value is the human bottleneck, not the use case: anything can be solved, the hard part is finding the problems worth solving. Ask each team to name the one piece of work that costs it most pain each week, pick the sharpest of those, and solve it properly with whichever single tool fits. The tool audit can wait; Gartner found only 13 per cent of organisations think they have the right AI governance in place, and governance follows focus, not the other way round.
How do I choose which AI use case to start with?
Not by scoring use cases on feasibility, because nearly everything is feasible now. Choose by bottleneck economics: a bottleneck is a cost, solving it is a return, and the return is weighed net, after the effort of solving. Rank the named bottlenecks by what keeping them costs in time, money or missed decisions, and start where the net return is largest and the team owns the timeline. A use case nobody bleeds for will be abandoned at the first friction, whatever its score on the matrix.
What is the Bottleneck Principle?
The Bottleneck Principle is a discipline inside the Havruta Methodology: the unit of AI value is the human bottleneck, not the AI use case. Anything can be solved; the hard part is finding the problems. In practice it inverts the usual sequence. Instead of cataloguing tools and hunting for uses, you name the points where human work actually jams, then bring the machine to the jam. It exists because use-case lists grow without limit while real bottlenecks are few, expensive and findable.
What is AI tool sprawl?
AI tool sprawl is the accumulation of overlapping AI tools, licences and now agents across an organisation faster than anyone can govern or properly use them. Gartner predicts an average global Fortune 500 enterprise will have over 150,000 AI agents in use by 2028, up from fewer than 15 in 2025, and found only 13 per cent of organisations think they have the right agent governance in place (Gartner, April 2026). Sprawl is a symptom: tools are being acquired on availability while adoption is not anchored to named problems.
Should we consolidate our AI tools?
Eventually, but not first. Consolidation before focus just produces a shorter list nobody uses properly. Sequence it the other way: anchor two or three teams to their named bottlenecks, see which tools those solutions actually need, and let the working stack reveal itself. Then retire what nothing depends on. Gartner's survey of 782 infrastructure and operations leaders found only 28 per cent of AI use cases fully succeed against ROI expectations (Gartner, April 2026); the failures are rarely caused by having the wrong tool and usually by having no anchored problem.
How many AI pilots should we run at once?
Fewer than you are running now, almost certainly. Grant Thornton's 2026 survey of 950 senior leaders found the organisations pulling ahead are scaling fewer pilots with better measurement and clearer exit criteria, and that those with fully integrated AI report far stronger revenue growth than perpetual piloters. A workable discipline: one bottleneck per team, a defined reading of success before the pilot starts, and a stated date on which the pilot either graduates or stops. A pilot with no exit criteria is not a pilot, it is a hobby.
References
- Gartner. "Gartner Identifies Six Steps to Manage AI Agent Sprawl." April 2026.
- Gartner. "AI Projects in I&O Stall Ahead of Meaningful ROI Returns." April 2026.
- Stanford Institute for Human-Centered AI. "The 2026 AI Index Report," Chapter 4: Economy. 2026.
- Grant Thornton. "2026 AI Impact Survey Report." 2026.