It is Monday morning. You have three things on your desk that need to move before lunch: a follow-up email to a prospect, a contract clause that needs interpreting before you sign, and a stack of customer support replies that are getting older by the hour. You could send the prospect email to your assistant. You could ask Claude to draft the support replies. You could pick up the contract yourself. Any of those round trips works, and any of them could be wrong.
The hard part is not the tasks. The hard part is the choice. Many owner-operators I meet have no clean instinct for which work goes where, so they default to “I’ll do it myself” and the day disappears under the weight of work that should have left their desk hours earlier.
There is a simpler way to make that call. Three variables, scored quickly, will settle the bulk of these decisions before you have even picked the task up.
What are the three variables that decide AI vs person vs you?
The three variables are confidentiality, judgement, and relationship. Confidentiality is who is allowed to see the inputs. Judgement is who is allowed to decide. Relationship is who the human counterpart is, and whose name needs to be on the output. Score each one as low, mixed, or high, and the right home for the task usually picks itself before you have spent any real time on it.
The reason this works is that those three variables capture the actual risk in delegating. Cost matters, but cost is downstream. If a task involves client personal data, the price of getting it wrong is an ICO investigation, not a higher hourly rate. If it depends on years of context you carry in your head, no agent can substitute for that, and a junior member of staff probably cannot either. If your name is going on the email, the relationship cost of a wrong tone is yours, not the agent’s.
The crude matrix is this. Low on all three: AI. High on confidentiality only: a person under contract. High on judgement only: you, possibly with AI as a thinking partner. High on relationship only: AI drafts, you sign. Two or three variables high: you, full stop. Hold that picture in your head and the day’s work sorts itself in seconds.
Why does this matter for your business?
Because the delegation choice itself has a cost, and many founders pay it badly. The UK virtual assistant market sits at £30 to £100 an hour. ChatGPT Plus and Claude Pro are £20 a month. For routine, low-confidentiality, objective work, the economics favour AI by roughly 17 to 1 against a £50-an-hour VA. That is real money for an owner-operator running a £1M to £10M services firm.
But the same maths flips the moment one of the three variables goes high. A task that touches personal data without anonymisation cannot go to a public LLM, regardless of price, because UK GDPR makes you the controller and the ICO has been increasing enforcement on technical and organisational measures. The ICO’s 2025 enforcement record shifted firmly toward security and processing failures, four of the largest fines ever issued by the regulator landed in that one year. A task that requires reading a client’s mood, or weighing a contract renewal against a relationship you have spent five years building, is not delegable to an agent that has not been in the room.
Where will you actually meet this in your week?
You will meet it on Monday morning in three flavours. The first is the obvious AI task: drafting a follow-up email, summarising a research document, reformatting a spreadsheet, generating a first pass at a job description. Low on every variable, send it to the model, edit the output, send it on. The second is the obvious person task: anything client-facing, anything with personal data, anything where someone needs to be on the contract.
The third is the grey zone, and this is where many founders lose time. Tasks where the right answer is AI-drafted, person-reviewed, founder-signed. A proposal where Claude generates three angles, your assistant tightens the one you choose, and you put your signature on it. A board paper where the agent pulls the data, your finance person sense-checks the numbers, and you write the recommendation. A reference letter where the model gives you a clean first pass and you make it sound like you. The grey zone is where the AI plus VA combination earns its keep, and the reason a lot of your week feels harder than it should is that you are still treating the choice as a binary.
When should you ignore the framework and just do it yourself?
When confidentiality and judgement are both high and the task carries reputational weight you cannot transfer. Pricing calls. Hiring decisions. Equity conversations. The first email to a client whose contract is in trouble. These belong on your desk because the cost of getting them wrong is paid in trust, and trust is not outsourceable. Delegating them to AI is a delay mechanism, delegating them to a junior is unfair.
The trap is the inverse: using “I’ll just do it” as the default for tasks that do not warrant the founder’s desk. Expense reconciliation. Calendar work. First-pass research. Internal documentation. Copy editing. These should leave your desk fastest, because every hour you spend on them is an hour the actually founder-shaped work is waiting for you. If a task does not need confidentiality, judgement, or your name on it, the answer is not your desk, regardless of how quick you think it would be to knock out yourself.
What are the related concepts to keep on the same shelf?
Two references will sharpen the framework. William Oncken’s “Who’s Got the Monkey”, in HBR, argues that accountability migrates upward unless you actively manage it. Apply that to AI: if you delegate to a model and the output goes wrong, the monkey is still on your desk because the model has no reputation at stake. Aaron Levie’s “every agent needs a box” framing adds the operational rule: agents need explicit scope and oversight points.
The third is the practical sibling to this post, when AI replaces a VA, which works through the same matrix on a single recurring example. Read it next if you want to see the framework applied end to end on the most common delegation question owner-operators ask. The pattern this post argues for, AI for the routine, person for the contracted and confidential, you for the trust-bearing, is closer to what the best delegators have always done. The shift is that the tools now hold a defensible first pass at the kinds of work an extra pair of hands used to take, which makes the threshold for delegating cheaper to cross. Score the three variables, pick the channel, get the day back.
