Technology markets often move from breakthrough to extravagant promise, disappointment and more routine use. Gartner's Hype Cycle popularised one version of that pattern. It is a lens, rather than a timetable or a law.

The danger for a professional services firm is treating “AI” as one position on one curve. Document retrieval, client communication, forecasting and autonomous decision-making have different evidence, harms, dependencies and maturity. A persuasive demonstration of one task says little about another.

When the noise fades, durable advantage will come from organisations that can select, evaluate, govern and stop specific uses. The capability is operational judgement, not enthusiasm or cynicism about AI as a whole.

Replace cycle predictions with task evidence

The source declares autonomous agents and creative work to be at particular stages by particular dates. It also predicts which applications will be productive in 2028. Those claims are unstable and unsupported.

Instead of asking where AI sits on a curve, take one proposed use and ask:

  • Who is doing the task, and what problem is visible in the current process?
  • What evidence would show that AI improved quality, time or cost?
  • Which information and systems are required?
  • What could go materially wrong, and who can reject the output?
  • What result would justify expansion, redesign or stopping?

Narrow uses can still create wide consequences. Summarising a document may sound contained, yet omitted qualifications can affect advice. Drafting a proposal may save time while weakening the firm's voice or disclosing confidential examples.

Adoption without a use is a gym membership

The source describes a 200-person consultancy buying an enterprise licence, holding two days of training and then pausing because returns did not appear. The article compares this with attending a gym induction and wondering why fitness did not improve.

The case and quotation require verification. Keep the analogy. A licence and general training create access; they do not redesign a workflow, establish quality or allocate responsibility.

Distinction also admits that its own early content pilot was framed as “explore how AI can improve our content process” and produced activity without a useful outcome. That is exactly the kind of experience the refined instructions say to preserve. It shows a change in practice rather than an immaculate case study. Verify the internal record, then retain the lesson: specific problem, measure and owner before tool enthusiasm.

Durable uses have an operating home

An AI-assisted workflow is more likely to survive when four conditions hold.

The task and value are bounded. The team knows the input, output, current performance and person affected. Success involves quality and outcome as well as speed.

Governance exists before live use. Data, supplier, permissions, review, records, incidents and stop authority are designed into the workflow. Requirements vary with sector, client contract and consequence.

The use fits systems and work. People can access approved information, use the output at the right point and recover when the tool fails. A separate impressive interface may add work overall.

The organisation learns from use. Qualified people review difficult cases, corrections and near misses. The team updates evaluation rather than relying on initial acceptance tests.

These conditions sound unglamorous because they are the work hidden by a demo. They also determine whether a pilot becomes a supported service.

Claims that age badly

The source identifies four categories worth auditing now.

Strategy without commitments

An AI strategy stored after board approval becomes an artefact of the moment. It needs named work, decision gates, budget and a review rhythm. Distinction's separate article on what an AI strategy should contain develops that operating model.

Marketing beyond actual capability

“AI-powered insight” can imply more than occasional use of a public drafting tool. Claims should identify the capability and preserve the responsible professional's role. Legal, regulatory and advertising review may be required.

Avoid positioning the absence of AI as backward. A firm can make a responsible decision that a task does not benefit from it.

Contracts signed around hope

Review multi-year commitments, minimum usage, data rights, model changes, service levels, audit evidence, subcontractors, price, export and termination. A supplier that was appropriate for a pilot may not be supportable at scale.

The source's claim that vendors had exceptional leverage during a fixed historical window is unevidenced. The procurement lesson stands without it: buy against a use and exit plan, rather than a category narrative.

Governance added after data has moved

The source reports two incidents involving confidential data and costly remediation at professional and financial-services firms. These are high-risk client claims and must be verified with Legal, Security, Privacy and client permission. Do not use them as casual proof.

Before a pilot, classify information, approve the supplier and environment, limit access and define retention. If the use cannot operate within those boundaries, stop.

Waiting and rushing are both portfolio decisions

“Wait until things settle” can avoid weak suppliers and premature investment. It can also leave the firm with no internal evidence or capability. The right answer differs by task.

Use a portfolio:

  • Proceed with a low-consequence, measurable investigation
  • Prepare data or governance without deploying
  • Monitor a capability until evidence or supplier conditions improve
  • Reject a use whose harms or economics are unacceptable

This avoids a false choice between an enterprise AI programme and inactivity. Readiness can improve through data ownership, process mapping, evaluation design and staff guidance even when no live AI use is approved.

The source describes the downturn as the best counter-cyclical investment moment and borrows financial-market logic. Technology adoption does not offer the same return mechanism, and competitors pulling back is weak reason to proceed. Lower noise may improve supplier conversations; only task evidence should justify investment.

Run a controlled comparison

Choose a recurring task with a suitable owner and enough examples to test. Build a representative set including difficult and high-consequence cases. Compare current work with the assisted workflow.

Measure:

  • Accuracy, completeness and material error
  • Time across the complete task
  • Review, correction and rework
  • Cost, including qualified people
  • User and affected-person experience
  • Privacy, security and professional incidents
  • Variation across cases

Set pause conditions in advance. Preserve human routes and records. Ninety days may suit one use and be meaningless for another; choose duration from task frequency and evidence needs.

If the use works, expand cautiously and retest under new volume, users and integrations. If it fails, record why. A stopped pilot can create more institutional capability than a nominally successful deployment nobody critically examined.

Capability is the residue after hype

The source argues that learning-by-doing creates institutional intelligence that cannot be bought. That is directionally right when the experience is structured and safe. Repeated casual use can also embed weak habits.

Develop people who can frame problems, question outputs, understand data boundaries, evaluate suppliers and connect technical performance with client or professional consequence. General awareness training is only the start.

The firms that will be left standing aren't the ones that made the most noise during the hype. They're the ones that used the noise as cover to do the real work.

If you want to benchmark where you actually stand before deciding where to start, our AI readiness checklist for mid-market service firms is a sensible first step.