Set the business result before you set the budget: name the number that has to move, the date it has to move by, and the person accountable for it, then let the tool shortlist follow from that.
Hype is expensive in one specific way. It moves the purchase ahead of the problem.
The hype cycle is a purchasing problem
Money leaves the business at the point of purchase, long before anyone can measure whether the tool works. That is where the cycle does its damage.
MIT Sloan reported on Akamai chief technology officer Robert Blumofe setting out the sequence he sees repeatedly: an early AI success becomes theatre, theatre produces fear of missing out, and fear produces a failed implementation. The MIT Sloan account of his EmTech AI talk is worth reading in full, because the failure he describes is not technical. It is a failure of sequencing.
The pattern has a track record. The Conversation noted that roughly 90% of enterprise blockchain solutions had failed by mid-2019, and it draws a straight line from that period to this one. The blockchain comparison also covers the Klarna chatbot reversal and CNET's error-strewn AI articles. None of those projects failed because the technology was incapable. They failed because the problem was never specified tightly enough to fail honestly.
Buyers push in the same direction as the vendors. UK Tech News reported that 93% of service and support technology buyers rated AI capability as important to their purchasing decisions, citing the 2025 Gartner Business Buyer Survey. That figure describes demand. A capability checklist can be satisfied by a product that solves nothing.
Wildnet Edge's guide to the Gartner Hype Cycle for 2026 is useful for one reason: it shows that the items on a typical AI shortlist are not at the same maturity. Agentic AI sits near the peak. Generative AI has slid into the trough. Treating both as one budget line guarantees that part of the money is spent at the worst possible moment.
Write the outcome as a number and a deadline
Nasscom's contributor, drawing on two years of work with chief information officers and chief data officers, argues that getting an agent to function is rarely the hard part; the difficulty is "what happens after deployment". That sentence belongs on the wall of every steering committee.
Underwriting shows the difference. Cutting a 30 to 40 day policy process in half is a workflow tweak, and it will be priced like one. Asking whether a policy could be underwritten in hours is a different question, because it forces the whole process to be rebuilt. The first target buys automation. The second buys a redesign, and only the second produces a step change.
Write the target in one sentence and make it survive a challenge: the metric, the current baseline, the target value, the date, the named owner, and the condition under which you stop spending. If the sentence cannot be written, the project is not ready for a vendor meeting.
MIT Sloan reported that Blumofe wants the burden of proof reversed. The question is not why a human should keep a task instead of AI, but what problem is being solved and which technology suits it. An outcome that cannot be expressed as a number is a preference, and preferences should not draw capital.
Rebuild the process before you pick the tool
Automation bolted onto a broken process makes the breakage faster. Nasscom's contributor makes the same point from the operating model side: the organisations that see step changes redesign how work is done, while the ones that retrofit AI into structures built for a different era stay stuck in pilots.
Run the sequence in this order.
- State the target as a number, a date and an owner.
- Map the process end to end and mark every handoff where one person waits for another.
- Delete or merge steps before any system touches them, because several handoffs exist only to move documents between desks.
- Choose the smallest technology that covers what remains, and prefer a purpose-built model where the task is narrow.
- Fix a module boundary so the model can be replaced without rebuilding the workflow around it.
Point five is not housekeeping. The Centre for Future Generations records that AI model prices fell by 99% in 18 months, a cost decline that took solar power roughly 40 years to achieve, in its briefing beyond the AI hype. Cheaper capability is arriving on a short cycle. A process welded to one model or one vendor is a bet against that curve.
Context is the foundation your data plan misses
An agent's recommendation is only as good as the history behind it. Data volume does not buy trust; context does.
Nasscom's contributor sets this out in underwriting terms. An experienced underwriter is valuable because of pattern recognition across similar clients, similar brokers and similar risk profiles, and most of that knowledge is written down nowhere. A building at a given address has prior tenants, prior claims and a record of how comparable risks were written. Feed an agent that history and it can produce something a professional will act on. Feed it only the current submission and it is guessing with confidence.
The numbers support the worry. Nasscom cites the EXL 2026 Enterprise AI Study, which found that only 30% of surveyed organisations report data available enterprise-wide, with data quality the most stubborn foundation problem. The lesson is not to fix the whole estate. It is to fix the slice of the value chain you are transforming, and to define that slice before the integration work starts.
Roles move at the same time. The assistant who prepared documents becomes the reviewer who handles exceptions and applies judgement, which is a harder job and needs deliberate reskilling, clear accountability and support. Insurance and healthcare also face the retirement of people whose expertise was never documented. That knowledge can be captured, but only if someone is given the job.
Put agent assurance in the operating model from day one
Assurance is a permanent function. Nasscom's contributor names it as a discipline most organisations do not yet have: continuous verification that an agent performs reliably, transaction by transaction, without drifting as it meets new documents, emails and data.
The risk profile is unfamiliar to software teams. A flawed agent decision can replicate across every transaction it touches, in the way a manufacturing batch defect does. That makes a kill switch mandatory, and it makes the testing estate part of the design: a simulation environment to exercise behaviour before production, and a live measurement environment covering reliability, compliance and cost.
Two additions from experience. First, the assurance lead should report into risk or internal audit, because a function that grades its own homework is a press release. Second, test the kill switch on a schedule with someone who did not build it. An untested stop button is a diagram.
Governance works better when it arrives early. Nasscom's contributor recommends bringing finance, compliance and security into a shared council before a problem surfaces, so that risk is designed in.
Score every candidate against four kinds of value
Score the shortlist on value the buyer can bank. The Pragmatic Institute teaches a four-lens test that does this well, and it takes an afternoon.
Buyer value asks whether the person who signs the cheque can see the return. Score it on cost reduction, revenue growth, scalability, risk mitigation, differentiation and speed to measurable benefit.
User value asks whether the daily work gets better. Score it on time saved, accuracy, friction removed, personalisation, accessibility and retention over time.
Vision value asks whether the investment supports where the business is going. Score it on strategic alignment, readiness for regulatory or market shifts, fit with the existing technology estate, credibility in the market and the ability to attract people who want to build with it.
Competitor value asks whether the position improves. Score it on uniqueness, parity with rivals where parity matters, first-mover advantage, customer loyalty, operating cost and the speed of future product cycles.
A tool that scores high on vision and zero on buyer and user value is a research project. Fund it from an innovation budget and say so out loud. That single distinction kills most hype purchases before they reach procurement.
Nasscom's contributor closes on the same discipline: begin with a small number of well-defined agents, build governance and assurance alongside them, upskill the people who hold the domain knowledge, and expand only where the evidence supports it. The firms that get value from AI are rarely the ones with the newest model. They are the ones that can point at a process and describe exactly what changed.
This page draws on the Nasscom community contribution, Why AI transformation must start with business outcomes.
Sources
- How to break the 'AI hype cycle' and make good AI decisions for your organization
- The Gartner Hype Cycle: A Smart Framework for AI Investment in 2026
- Stop Chasing AI Hype and Start Building Strategy
- The AI hype is just like the blockchain frenzy – here's what happens when the hype dies
- Stop Chasing AI Hype: Start Solving Market Problems
- Beyond the AI hype – the FAQ
- Why AI transformation must start with business outcomes
Source: Why AI transformation must start with business outcomes, Nasscom





