B2B AI PlatformB2B AI Platform
Why Ops Expertise is Your AI Secret Weapon
B2B AI Platform

Why Ops Expertise is Your AI Secret Weapon

Emma RobertsWords by Emma Roberts

Two years ago the fight inside most businesses was about permission to run an AI pilot. The fight now is about who absorbs the cost when the pilot never becomes an operation. The authorisation costs little. Recovery is where the money is spent. TechSpective's September 2026 report, From Innovation to Impact, is the article this page covers, and it reduces the problem to one line worth carrying into a steering meeting: "the challenge is deployment."

That line is a management problem appearing as a technology issue. The people who have run the queue, the depot, the collections desk or the service centre decide whether an AI programme pays back. What follows explains why, and what has to be in place before they can.

Where AI projects actually break

Most AI projects that fail are projects whose software works.

TechSpective reports that McKinsey's latest global survey states AI use is at 78% of organisations in at least one business function, up from 55% two years earlier. The same article records estimates that most AI initiatives miss their intended outcomes, and that they fail more often than ordinary IT projects do.

Adoption climbed. Delivery did not. No part of the model explains that gap, because the model is the same in every one of those cases and the variable is in another place: where the model meets the process. This is the step that no one documented. The exception that gets fixed by phone. The customer who must never be routed to a bot.

TechSpective reports that the organisations extracting the most value from AI are significantly more likely to prioritise business process transformation alongside the technology itself. Treat that as an order of operations. The process work comes first, and the tool fits into it.

Knowledge that lives inside the operation

The late-shift supervisor knows which complaint is about to escalate, and that knowledge sits in no training set.

TechSpective's example is a customer service assistant trained on past interactions. It recommends a plausible reply and still misses the escalation trigger, the shift in tone, or the regional preference an experienced adviser spots in two seconds. The remedy the article proposes is procedural: frontline teams join model design and tuning as authors of the rules the assistant runs on, from the first draft. Adoption and customer outcomes both move when that happens.

This is a data question before it is a culture question. An operation runs on tribal knowledge, and tribal knowledge costs what it costs precisely because it is unwritten and unevenly spread. A model trained on the log inherits the average of what happened. The judgement of what should have happened stays with the people who did the work.

Write the procedure down before you automate it

The first real deliverable of an AI programme is an explicit procedure with an owner.

smartbot.cloud's guide to support and ops assistants makes the distinction cleanly. The assistants that remain effective are productised workflows built from approved knowledge. A simulation of a trusted human is a different product with a different failure mode. Policy answers, troubleshooting trees, fulfilment steps, access requests and escalation rules each have a source and a limit. The guide's instruction is to define what the assistant must never do before defining what it should do, and to keep the tiers of read, recommend and act separate.

Decagon explains the same idea from the platform side. Its agent operating procedures take the standard operating procedures a service team already has and compile them into validated workflows an agent can execute, such as processing a refund or verifying identity. Natural language stays with the operational team. Guardrails, integrations and security stay with engineers. Business experts shape the behaviour and engineers manage the control plane.

The configuration limit is short. It should be.

assistant: refunds-triage
may_read: order_history, ticket_log, policy_cards
may_recommend: refund_eligibility
may_act: create_ticket, request_approval
may_not: issue_refund, change_payment_details
escalate_when: confidence_below_threshold, customer_requests_human
owner: service_desk_lead
review: monthly

Every line is a decision an operator has already made a thousand times. Writing it down exposes the disagreement: two supervisors, two escalation rules, and no way to test either one until now.

What the live queue exposes

An operation that cannot say which alert matters cannot delegate any of them to a model.

Kong's explainer on AIOps states the average cost of IT downtime is roughly $9,000 per minute and says teams receive upwards of 10,000 alerts a day, 99% of them noise. Point a model at that stream without operational judgement and it ranks the noise faster. The fix is unglamorous and entirely operational: correlate related events into a single incident, surface the root cause, and let routine work run on rules.

The engineer who has been on the receiving end of a false alarm at 03:00 makes the judgement that decides what earns a page at 03:00. That judgement is the specification.

How the UK's most demanding operator is doing it

The Ministry of Defence is rebuilding how it works, and treating AI as part of that rebuild.

The MOD letter published by GOV.UK in June 2026, signed by the Defence Secretary alongside the service chiefs, states that realising the benefit of AI "will require changes to processes, culture, behaviours and more". It treats data foundations as a prerequisite: a federated Defence data mesh, accessible and discoverable data, personnel and financial sources joined up. It asks for modular, open-architecture procurement that avoids vendor lock-in, sandbox environments for ordinary users, training recognised in the skills framework, and a route from local experiment to capability at scale. That is an operations plan with AI inside it.

The cautionary half of the same story matters. POLITICO reported that the Ministry used a Palantir-built model to sift submissions to the Strategic Defence Review, and that a defence industry figure dismissed the likely output as "a glorified word cloud", with concerns that bidders would game the keyword filter and that human judgement was being removed from a task that needed it.

Kenneth Payne, writing for King's College London, sets out the counterweight. The Ministry needs a blend of technical and non-technical understanding, and the only reliable way to learn how people and machines behave together is repeated exercises, wargames and simulations. Operational expertise is how the technical work gets validated.

Responsibility remains after handover

The failure no one catches is the one that falls between two teams.

TechSpective identifies fragmentation as the quiet killer: responsibility split across consultants, technology providers, operations and the business, with no thread running from strategy through deployment to adoption. It cites MIT Sloan research showing stronger digital transformation outcomes where accountability is clear across the whole implementation, and it argues that the same experts who design a solution should stay with it through optimisation and long-term adoption.

Read that as a staffing rule. One name against each procedure in the register, carried from the problem statement through to the quarter after go-live. The person who signs the escalation rules is the person the service desk calls when those rules produce a wrong answer. If the same name sits against the refund triage procedure, the assistant configuration and the monthly review, the programme has an owner. If the name changes at handover, it has a supplier and a dispute waiting to happen.

The point is sequence. Operations first, the technology fitted to what they already do, and accountability that outlives the launch. A model is a commodity that improves on a timetable set by another party. The queue, the escalation rules and the person answerable for both belong to you, and they are what turn adoption into a number a finance director recognises.

Sources

Source: From Innovation to Impact: Why AI Success Depends on Operational Expertise, TechSpective

Changed

Recently filed

More on this

Aligning LLM Integration with Corporate Governance
Aligning LLM Integration with Corporate Governance
AI Architecture for Insurance Brokerage
AI Architecture for Insurance Brokerage
Recursive AI Validation: The Technical Risks
Recursive AI Validation: The Technical Risks