The AI Chief of Staff Playbook: How I Run a 5-Venture Portfolio with Autonomous Agents

Running five ventures does not create five neat sets of work. It creates one continuous stream of decisions: a buyer needs an answer, a price needs checking, a piece of content needs approval, a meeting creates six follow-ups, and an adviser needs the right context before a call.

The conventional answer is a larger central team. My answer has been different: build an operating architecture in which autonomous agents handle defined parts of the flow, while people retain judgement, relationships and accountability.

I call the centre of that architecture Salem, my AI Chief of Staff. Salem is not one all-knowing chatbot. It is a coordination layer across email, meetings and WhatsApp, supported by specialist agents inside each venture. Its job is to keep context moving, surface decisions and reduce the number of small operational gaps that become expensive later.

This is the practical playbook behind that system: what each layer does, how the pieces connect, where humans intervene, and how I decide whether an automation is producing a return.

The architecture: one coordinator, many narrow specialists

The most useful mental model is not “an AI employee”. It is a small operating system with three layers.

The first is signals: messages, meeting notes, CRM records, pricing inputs, content briefs and workflow events. The second is specialist agents with narrow jobs. The third is human control: approvals, exceptions and accountable owners.

Salem sits between the signals and specialists. It watches for work that requires a response, turns unstructured information into a clear task, routes that task to the right workflow, and brings the result back for review when judgement is required.

A typical day might begin with Salem summarising priority emails, extracting commitments from meeting notes and identifying WhatsApp conversations that need a commercial follow-up. It does not decide every answer. It decides what deserves attention, prepares the context and prevents the hand-off from disappearing.

That distinction matters. The value is not that an agent writes text quickly. The value is that the operating loop closes reliably.

Salem: the portfolio coordination layer

Salem works across three channels that carry most of my operational context.

Email

The email workflow classifies messages by commercial importance, recognises requests that need a factual response, and prepares a proposed next action. A supplier question and an inbound advisory enquiry should not enter the same queue. A buyer asking for a price should not be treated like a newsletter.

The agent can assemble context from earlier exchanges, identify missing details and draft a response. I remain in the loop for pricing commitments, sensitive relationships and anything that could create legal or reputational exposure.

Meetings

Meeting notes become useful only when they change what happens next. Salem turns transcripts or notes into decisions, owners, dates and unresolved questions. Follow-ups are drafted with the relevant context rather than a generic summary.

The benchmark here is not “minutes saved writing notes”. It is the percentage of agreed actions that reach an owner and the time between a meeting and the first meaningful follow-up.

WhatsApp

In food and agro trade, material conversations often happen in WhatsApp. The challenge is not access to the message; it is preserving context across a fast-moving relationship. Salem helps identify unanswered requests, capture commitments and prepare the next response without pretending that a relationship can be automated end to end.

WhatsApp is also where human-in-the-loop discipline matters most. Tone, timing and trust carry more weight than perfect prose.

Rasm.cc: eight agents, one brand system

Rasm.cc uses eight specialised agents across content planning, research, drafting, adaptation and community work. Their narrow roles are deliberate. A research agent should not quietly become a brand strategist, and a community agent should not invent a product claim because it needs a quick reply.

The agents share a common source of truth: brand positioning, audience definitions, approved claims, editorial examples, banned language and review rules. That is what I mean by brand DNA.

Prompt engineering begins there. A long prompt full of adjectives is not a brand system. The useful material is operational: examples of what good looks like, examples that should be rejected, a hierarchy of claims, rules for evidence, and a clear answer to who can approve what.

The result is not “AI-generated content”. It is a controlled production line in which specialist agents can create options at speed and a human can judge whether the work deserves to represent the brand.

OleaDesk: autonomous prospecting with commercial boundaries

OleaDesk focuses on the sales desk for food and agro trade. It can research prospects, qualify whether a buyer matches the offer, prepare personalised outreach and keep follow-up moving.

The important design decision is what the agent is not allowed to do. It does not invent availability, commit to a specification or negotiate a price outside an approved rule. Those are commercial decisions with consequences.

Instead, the system handles repeatable desk work: gathering context, ranking opportunities, drafting messages and maintaining the follow-up cadence. The human takes over when a qualified conversation requires trust, judgement or a binding commitment.

For ROI, I watch response time, qualified conversations per hundred prospects, the proportion of follow-ups completed on schedule and the amount of human time spent per qualified opportunity. Sending more messages is not a useful success measure if relevance falls.

Olea Engine: instant quoting connected to the CRM

Olea Engine addresses a different bottleneck: turning a buyer request into an accurate commercial response. The system combines product and pricing logic with an MCP-connected CRM, so an agent can work with current customer context rather than an isolated prompt.

An instant quote is only valuable when the inputs are controlled. Product, packaging, freight, margin, customer terms and validity all need explicit rules. The agent assembles the quote and records the context; exceptions route to a person.

This changes the operating benchmark. Instead of measuring whether someone can produce a quote, I measure quote turnaround time, error and exception rates, conversion from quote to qualified discussion, and the number of manual touches required before sending.

The technology is the least interesting part. The hard work is making the commercial rules legible enough for a system to follow.

Forward-deployed advisory at RCS UK

My work with RCS UK takes these lessons into public-sector and non-profit organisations. “Forward-deployed” means working close to the actual operation: understanding where information enters, who owns a decision, which constraints are real and what level of automation the organisation can responsibly absorb.

The goal is not to arrive with a generic AI strategy. It is to identify a bounded workflow, establish its baseline, design controls and build enough of the new operating model that leadership can evaluate evidence rather than a presentation.

In these environments, governance cannot be added at the end. Data access, auditability, accessibility, accountability and escalation need to be part of the architecture from the first workflow map.

Three lessons from running agents in real operations

1. Human-in-the-loop is a design decision

“Human in the loop” is often used as reassurance without specifying the loop. A real control names the decision, the person accountable, the evidence they see and the point at which work stops without approval.

I use three broad levels. Low-risk, reversible work can run automatically and be sampled. Material external communication is prepared by an agent and approved by a person. Pricing, contracts, sensitive data and reputational decisions remain explicitly human.

The control should match the consequence, not the novelty of the technology.

2. Prompt engineering starts with brand and operating truth

A strong prompt cannot compensate for an organisation that has never written down its rules. The best agent work starts by making tacit knowledge explicit: how a good lead is recognised, which claims are approved, how an exception is handled, what tone belongs to the brand and where the system must admit uncertainty.

This is why implementation begins with interviews, examples and workflow mapping rather than a model selection exercise. The prompt is the final expression of operational thinking, not the first step.

3. ROI needs a baseline before the build

An AI project should begin with a number that can move. Depending on the workflow, I establish a baseline for cycle time, response time, cost per completed unit, manual touches, exception rate, conversion rate or work-in-progress.

I then compare the new workflow against that baseline after allowing for review and correction time. Gross minutes “saved” by a model are not a return if a senior person spends those minutes fixing weak output.

Useful ROI benchmarks answer four questions:

1. Did the workflow become faster?
2. Did quality or conversion hold or improve?
3. Did the number of human interventions fall in the right places?
4. Did the system create new risk or hidden maintenance work?

The benchmark must survive contact with the whole process.

Where to begin

Do not begin by trying to build your own Salem. Begin with one operating constraint that recurs often enough to measure. Map the real workflow, including exceptions. Name the human owner. Establish the baseline. Then design the narrowest agent that can change the number without taking on decisions it should not make.

That is also how I approach client work. The Executive AI Diagnostic is a one-week operational and architecture audit. It identifies high-return opportunities across sales, marketing and operations, then delivers three custom agent blueprints with ROI models. The £500 fee is credited in full towards a deployment sprint.

Book the Executive AI Diagnostic if you want to turn one of your operational bottlenecks into a controlled, measurable agent workflow.