Building AI vs. Knowing What to Build: Where the Real Bottleneck Sits

A large financial institution has three hundred engineers and connections to every major model. Its transformation lead describes the difficult phase as discovery, because the organization sometimes does not know what it wants.

That combination is more common than the state of AI tooling would suggest. Build capacity has become abundant and comparatively cheap. Knowing which problem deserves the capacity has not.

The two constraints look similar from the outside, since both produce a programme that is busy and not delivering. They call for opposite investments. An organization that adds engineering capacity to a discovery constraint builds more of the wrong things faster.

Telling them apart takes one question, and most organizations have not asked it.

Key takeaways

• Build capacity and discovery capacity are distinct constraints with opposite remedies.
• The diagnostic question is whether the last five completed initiatives moved the metric they were meant to move.
• A backlog ranked by advocacy is a discovery constraint wearing the appearance of a prioritization problem.
• Organizations with strong engineering frequently have the weaker discovery capability, since the strength attracts the requests.
• Discovery capacity is cheaper to add and shorter to build than engineering capacity.
• Both constraints can be present, and the sequence matters: discovery first.

The two constraints

The bottom row is the asymmetry that matters. Under-investing in build slows down value. Under-investing in discovery produces velocity in the wrong direction, which is more expensive because the output persists and has to be maintained.

The diagnostic question

Look at the last five initiatives your AI programme completed. For each one, name the business metric it was meant to move and whether it moved.

If most shipped and most moved their metric, the constraint is build capacity. The selection process is working and the limit is throughput.

If most shipped and few moved their metric, the constraint is discovery. The organization can build and is building the wrong things, which no amount of additional capacity corrects.

If few shipped at all, the constraint is build capacity or governance, and the discovery question is premature.

If nobody can name the metric for most of them, the constraint is discovery and it is upstream of where anyone has been looking. An initiative without a named metric was never selected against a problem.

That last case is the most common in large organizations and the easiest to mistake for a measurement gap. It is not a measurement gap. A metric that was never named could not have been used to choose the initiative.

Why strong engineering organizations develop weak discovery

There is a mechanism here worth naming, because it makes the pattern predictable rather than surprising.

An organization that builds well attracts requests. Business units learn that if they arrive with a specification, something gets built. Over time the intake becomes a queue of specifications rather than a queue of problems.

Specifications are easier to prioritize than problems, because they can be compared on effort. So the prioritization process optimizes for what it can compare, and the question of whether each specification addresses a real constraint stops being asked.

The result is an organization with excellent delivery metrics and a portfolio disconnected from operational reality. Nothing in that sequence involves anyone making a mistake.

Bottleneck diagnostic

Reading the last five initiatives

The second and fourth rows account for most enterprise AI programmes that are busy without delivering.

What a discovery constraint actually costs

The cost is not the wasted build, which is usually recoverable. Three costs compound beyond it.

Maintenance. Every shipped initiative has to be maintained regardless of whether it delivered value. A portfolio of systems addressing steps that were not constraints consumes engineering capacity indefinitely.

Credibility. Business units that receive several systems without operational improvement stop bringing real problems, because the pattern suggests the exercise does not help them. The intake degrades further.

Opportunity. The constraint that was not addressed is still there, compounding. In processes where a handoff or an approval governs throughput, every quarter it stays unaddressed is a quarter of the same cost.

The third is the largest and the hardest to put in a business case, because it requires knowing what the constraint was, which is the thing the organization does not have.

Why discovery capacity is the cheaper investment

Engineering capacity is expensive to add, slow to onboard and difficult to reduce. Discovery capacity has a different profile.

It is faster to establish, since it requires access to the people running the processes rather than a hiring cycle. It is cheaper per unit of decision improved. And it is reusable, since a current-state picture informs many initiatives rather than one.

There is also a sequencing argument. Discovery reduces the amount of build required, by identifying steps that should be removed, integrated or standardized rather than built around. An organization that runs discovery first frequently finds its build backlog smaller than it was.

When both constraints are present

Frequently they are, and the sequence is not symmetric.

Running discovery while build capacity is saturated produces a validated backlog that nothing can act on, which is frustrating but recoverable. Adding build capacity while the discovery constraint is unresolved produces more systems addressing non-constraints, which adds maintenance load and makes the portfolio harder to correct.

The practical sequence is discovery first, at a scope small enough to complete while the delivery queue clears. The output arrives as the delivery capacity becomes available.

Where Horizon fits

Horizon is an AI-powered continuous discovery platform. It addresses the second constraint, which is establishing what deserves to be built.

Discovery Cycles run AI-led interviews across the roles that operate a process, asynchronously and at census scale, adapting to each role and following up on gaps. The Insights Dashboard ranks findings by impact and effort with traceability to the input behind each one, which is what turns a request into a scoreable candidate. The Initiatives Dashboard converts priorities into business cases with owners, effort and expected impact attached, which is the artifact a build decision requires.

La Segunda, one of Argentina's main insurance groups with more than 90 years of history and 1,200 offices, illustrates the pattern at the level of one process.

Its long-term injury claim process ran across several systems, including three named platforms plus spreadsheets, with little integration and more than 700 active cases handled with high effort and limited visibility. The systems existed. What did not exist was an account of where the effort concentrated and which improvement would matter most.

Horizon interviewed seven case managers and two medical auditors, covering 100% of the team. The full process was mapped and analyzed in 48 hours. Within five weeks the engagement had produced a complete BPMN process map, a dashboard of insights grouped by effort and impact, and 10 key findings covering automation, integration and workflow redesign opportunities. Conversation satisfaction was 8.2 out of 10, with 90% saying they would talk again, and the engagement saved more than 30 discovery hours.

The results fed directly into building a new medical follow-up platform for chronic case management, with Horizon dashboards guiding next steps on automation, alerts and system integrations.

The sequence is the point. The build decision came after the current state was established, which meant the platform was scoped against where the effort actually sat rather than against where it was assumed to sit.

That is one engagement under specific conditions rather than a projection for any organization.

Diagnostic checklist

1. For your last five completed initiatives, can you name the business metric each was meant to move?
2. For each, did it move?
3. Is your backlog made of problems or specifications?
4. When a request arrives, does anyone ask what happens today and what happens if nothing changes?
5. How many of your active initiatives have a named business owner accountable for the outcome?
6. What proportion of engineering capacity maintains systems that did not deliver their intended outcome?
7. If you doubled build capacity tomorrow, what would you build?
8. Can you state where the throughput constraint sits in your three highest-priority processes?

Question 7 is the fastest version of the diagnostic. If the answer is a longer version of the current backlog, the constraint is not capacity.

FAQ

How do you know if your AI bottleneck is building or knowing what to build?

Look at the last five completed initiatives and check whether each moved the business metric it was meant to move. If most shipped and few moved their metric, the constraint is discovery. If most shipped and most moved their metric, the constraint is build capacity. If nobody can name the metric, the constraint is upstream of both.

Why do companies with strong engineering teams still stall on AI?

Because build capability attracts specifications rather than problems. Business units learn that arriving with a specification produces a system, and the intake becomes a queue of solutions. Prioritization then optimizes for what it can compare, which is effort, and the question of whether each item addresses a real constraint stops being asked.

Should you invest in discovery or delivery capacity first?

Discovery, when both constraints are present. Running discovery while delivery is saturated produces a validated backlog that waits, which is recoverable. Adding delivery capacity while discovery is unresolved produces more systems addressing non-constraints, which adds permanent maintenance load.

What does a discovery constraint cost?

Beyond the wasted build, three things compound: maintenance of systems that did not deliver value, credibility with business units that stop bringing real problems, and the unaddressed constraint itself continuing to cost every quarter it stays in place.

Can a company have both constraints at once?

Frequently. The distinguishing question is which one to address first, and the sequence is not symmetric. Discovery reduces the amount of build required by identifying steps that should be removed, integrated or standardized, so running it first often shrinks the backlog rather than adding to it.

Capacity is not the scarce thing anymore

The ability to build has become abundant in most large enterprises. Model access is a procurement decision, engineering capacity is a budget decision, and integration patterns are well understood.

What stayed scarce is a reliable account of which problem is worth solving, and that is not something more capacity produces.

See it. Fix it. Lead it.

See Horizon in action.