Notes
Deployed is not operational
Using AI is common; running on it is rare. Five questions tell a pilot from an operational process, and four levels show the way up.
By Alessandro Singh and Rohith Varma, Updated , 5-minute read
In short
- A working pilot becomes a process when it has an owner, a measure the business already tracks and a fallback. Until then, it is still a pilot.
- McKinsey's 2026 survey shows the gap: AI is widely used, but most companies cannot yet see it in their profit and loss account.
- Move each process up the four levels one at a time, from assist to run and improve, and let it earn each step.
- For high-risk uses listed in the AI Act, such as recruitment and credit scoring, oversight, monitoring and logs become legal duties from 2 December 2027.
Wide use, little effect on the numbers
Most companies we meet use AI somewhere. A chat assistant drafts emails, a pilot summarises tickets, a team has a demo that impresses the board. Few run any part of the business on it.
McKinsey's 2026 survey shows the gap. Eight in ten respondents said AI had improved their individual productivity. Only 37% attributed any EBIT impact to it, and about 6% were high performers, attributing 5% or more of EBIT to AI and calling the value significant. The gain is real at the desk, but most companies cannot yet see it in the profit and loss account.Source 1
In 2025 Gartner predicted that more than 40% of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls. Those are the causes we see. The model is rarely the problem. The work around it is.Source 3
Five questions that tell a pilot from a process
We ask five questions of every AI initiative. If the answers are missing, you have a pilot, however well it demos.
- Who owns it? In a pilot, the team that built it. In an operational process, a named person in the business, with time set aside.
- What does it measure? A pilot has a good demo. A process moves a number the department already tracks, reviewed every month.
- What happens when it fails? In a pilot, someone notices, eventually. In a process, monitoring raises it, and a fallback keeps the work moving.
- Who can change it? In a pilot, whoever has access. In a process, changes are reviewed, tested and logged.
- What does it cost to run? A pilot's cost is unknown until the invoice. A process is budgeted and compared with the value it creates.
If you cannot name its owner, its measure and its fallback, it is still a pilot.
Four levels, earned one at a time
Each process earns its autonomy one level at a time:
- Assist (deployed AI): AI drafts, finds and summarises. People decide and act.
- Automate with approval (specialised AI): AI completes steps. A person approves before anything is sent, posted or changed.
- Run with exceptions (operational AI): AI runs the process. Exceptions and edge cases go to people.
- Run and improve (operational AI): AI runs the process and proposes improvements. People approve each change.
Treat each process like a new hire: a job description, a sign-off limit, a manager and a monthly review. Some processes should stay at the second level, and some decisions stay with people: pricing, contract terms and anything decided about a person.
Before a process runs on its own, write your answers down as six controls. Three come from the questions: the owner, the measure and the fallback. The others are limits (the amounts, accounts or topics that always go to a person), a log of every step and decision with its evidence, and a stop: one switch that pauses it.
In the finance walkthrough (on closeweave.ai) on closeweave.ai, invoices are matched against the order and the goods receipt within a tolerance the controller sets. The AI never releases a payment. Every decision is logged with its evidence, held invoices queue for the team, and one switch pauses matching.
The law asks for the same controls
In the UK, Articles 22A to 22D of the UK GDPR, in force since 5 February 2026, set the rules for significant decisions made solely by automated means. The safeguards include telling people, letting them make representations, a route to human review and a right to contest.Source 5
The EU AI Act classes some uses as high-risk, such as recruitment and credit scoring. For those, it turns these controls into duties for the deployer, the company that uses the system. Deployers must:Source 6
- use the system according to its instructions;
- assign human oversight to people with the necessary competence, training and authority;
- monitor its operation, and if it may present a risk, suspend it and tell the provider and the authority;
- keep the logs under their control for at least six months.Source 6
The Digital Omnibus on AI, in force since 27 July 2026, moved the start of these high-risk duties to 2 December 2027.Source 7
An owner with authority, monitoring, a log and a way to stop: a well-run process has them anyway. We build them in from the start, alongside your legal or data protection lead.
Where to start
Pick one process that pays back, with a number the department already tracks. Answer the five questions in writing. Run it with approval until the measure holds, then let it run with exceptions. Then pick the next one.
How we help
Client resultFounder advisory, 2026
10x ARR: $200K to $2M in annual recurring revenue by November 2026 for a B2B professional-services founder. As their go-to-market team, we rebuilt pricing and qualification, provided market access, set up AI systems that gave back 20+ hours a week and closed the sales together with the founder. We also helped them raise their seed round.
The AI systems are the work of CloseWeave AI: we build AI into your systems, train your people to run it and hand it over, with an owner, limits and a log for every process.
More notes
AI you can rely on
Relying on a vendor's AI: the evidence to ask for
In a 2026 survey, vendors supplied over half of UK financial firms' most material AI. Firms still answer for it: ask for evidence that repeats.
For enterprises, investors and founders
Updated
, 8-minute read
Market access
Selling into the Gulf: Saudi Arabia and the UAE
Gulf buyers score what you build in the country. In Saudi Arabia and the UAE, that shapes your entity, your partners and your price.
For founders and enterprises
, 6-minute read
Newsletter
One email a month
Our newest notes, and what we are seeing in AI, machines at work and deep technology.
Sources
In the order we cite them, with the day we read each one.
1. The state of AI (opens in a new tab)
McKinsey & Company (QuantumBlack), , Accessed , Back to the text, where source 1 is cited
2. Companies' financial value from AI holds firm in 2026 (opens in a new tab)
FM Magazine, reporting McKinsey's The state of AI in 2026, , Accessed , Back to the text, where source 2 is cited
4. Over 40% of agentic AI projects will be scrapped by 2027, Gartner says (opens in a new tab)
Reuters, via The Star, , Accessed , Back to the text, where source 4 is cited
7. Regulation (EU) 2026/1744 (Digital Omnibus on AI) (opens in a new tab)
EUR-Lex, Official Journal of the European Union, , Accessed , Back to the text, where source 7 is cited