Running the business on the AI we sell
A demonstration never shows you where AI actually breaks. Running a real business on it, day to day, teaches you what no pitch does: what genuinely works, and what it costs to find out.
The clearest thing about running a business on AI is that a demonstration never shows you where it breaks. At Webjects, day-to-day operations, including client correspondence, scheduling, invoicing and quality checks, run through CLU, an AI Chief of Staff, working alongside a team of specialist AI agents and three human staff. That’s not a marketing claim. It’s a plain description of how our own back office works, disclosed in full because clients ask, and because it’s the most honest answer we can give to the question of whether AI can actually be trusted with a business.
What the team actually looks like
The team behind CLU is named publicly on our own website, not described in the abstract: named roles spanning account management and development through to finance, quality review and legal compliance, every one an AI agent working under human oversight rather than as a novelty. Three humans sit alongside them. That level of specificity is deliberate. This isn’t a metaphor for “we use AI a bit”, it’s closer to the literal org chart, and being able to point at it is the only reason anything below is worth reading.
What works
What works is anything repeatable with a clear right answer. Drafting a first version of a reply. Checking a set of figures against a source document. Running a scheduled report. Flagging a website that’s stopped loading properly. These are jobs with a defined shape, and a system doing the same defined job for the four-hundredth time is exactly as careful as it was on the first, which isn’t something most people can say of themselves by four o’clock on a Friday. That’s the genuine win: not intelligence, but consistency on work that used to lose people’s attention halfway through.
The part that actually took the work
None of it runs unsupervised, and that’s the part that took real effort to build. Every client-facing message goes through a review step before it’s sent. Work heading to a client is checked by a dedicated reviewer role first, and final sign-off sits with a person before anything reaches a client. That structure exists because an AI system will occasionally produce something confident and wrong, and the only real defence against that is a human checkpoint placed before a mistake can reach anyone outside the business, not after.
Building that structure took most of the real time, not the automation itself. Most of the effort went into writing down what the system got wrong and turning that into a rule it wouldn’t break again, over and over, rather than into building the system in the first place. That’s the unglamorous part most AI pitches skip.
What genuinely doesn’t work yet
What doesn’t work yet is judgement calls with no clean right answer: how to phrase something difficult to a client who’s upset, whether a design choice actually looks right rather than just measures right, when a polite no is the correct answer to a request. These stay with a person. So does anything that can’t be undone easily: a payment, a live change to a client’s website, a decision that touches someone’s account. All of it sits behind an explicit approval step rather than running automatically, however routine it’s become.
The lesson that surprised us
The lesson that surprised us most is that trust in a system like this isn’t built once, at setup. It’s rebuilt constantly, through the same small correction loop you’d use with a new member of staff: something goes slightly wrong, the reason gets written down, the same mistake doesn’t happen twice. A framework like this holds up because every error becomes a rule, not because it arrived error-free. Anyone expecting to switch AI on and stop checking it has misunderstood what it actually is.
What this means if you’re weighing it up for your own business
For another business weighing this up, the practical takeaway isn’t “buy the tool”. It’s to expect an ongoing relationship with whatever you introduce, not a one-off purchase, and to decide in advance what it’s never allowed to do without a person looking first. Start with the boring, repeatable, low-stakes work: the invoicing checks, the draft replies, the scheduled reports. Not the client-facing decisions. Build the habit of correcting it in-house, so the lesson doesn’t get learned the hard way in front of a customer.
Running a business this way hasn’t made AI feel more impressive to us. If anything, it’s made us more specific about where it’s actually useful, and considerably more careful about where it isn’t, which is exactly why we only add it to a client’s own website when there’s a real question it can answer well, not because it’s available. That’s a smaller claim than most AI pitches make. It also happens to be the honest one.
Sources
- How We Work (Webjects’ own live page, including the “Meet the team” roster). Read directly for the AI/human structure and the operating principle that AI is only added where it earns its place.
- llms.txt (Webjects’ public AI-disclosure file). Read directly for the named human and AI team roster, role titles, and CLU’s role as Chief of Staff.
- Internal, verified against source:
src/data/team.tsandsrc/data/services.tsin thewebjects-websiterepository. Team roles and the AI & Automation service description were checked against these files; no figures invented.