Own the brain. Rent the engine.
Software is getting cheap. Your data and judgement are not. Tools that took a vendor years to build can now be built in weeks. So the lasting value isn't in the software. It's in your years of candidates, clients and placements, and in how your best people decide.
01 / Where agencies are today
Every tool is adding AI, on its own terms.
Your CRM, your job boards, your email and your phone system all have an AI feature now, or will soon. Each one works from its own slice of your data, is priced per seat, and follows a roadmap set for thousands of agencies rather than yours.
You end up with five AI features in five tools that don't talk to each other, and no way to tell which of them is paying for itself.
Alex Lockey, Bench
This is not new. Recruitment has been renting its own working day for years. SaaS businesses became gatekeepers to productivity, especially in recruitment, locking businesses into expensive, multi-tenanted, modular, multi-year agreements. Every product was built for the median user, which meant workarounds for workflows the product team never considered.
Meanwhile the team does what people always do when a tool helps. Consultants paste CVs and client notes into personal AI accounts, with no record and no agreement. The agency's most valuable material is already leaving, one chat at a time. In Microsoft and LinkedIn's 2024 research, 78% of people using AI at work were bringing their own tools.
02 / The missing why
Your CRM stores what happened. It does not store why.
The last generation of enterprise software became valuable by being systems of record. Salesforce for customers. Workday for employees. SAP for operations. Own the data, own the workflow, own the lock-in.
A CV was sent. An interview was booked. A placement was made. What the record doesn't hold is why any of those things happened. The judgment. The context. In recruitment, that missing layer lives in recruiter notebooks, email threads, WhatsApp conversations, LinkedIn DMs, and mental models built over years of working a specific market.
It walks out the door every time a good recruiter leaves your business.
Alex Lockey, Why AI Can't Replace Your Best Recruiter (Yet)
Foundation Capital call the record of that reasoning a context graph, built from decision traces. For an agency, four things have to be true before it exists:
- Captured at the point of action. Not "CV sent", but "CV sent because of this context, this relationship history and this market timing", recorded as part of the work.
- Synthesised across systems. The reasoning pulls from the CRM, the inbox, messages, call notes and market knowledge at once.
- Exceptions made explicit. Why a stage was skipped, or a candidate sent to a client who isn't on the target list. These exceptions are where the real value lives, and they're the things nobody logs.
- Precedent made searchable. What did we do last time in a similar situation, and did it work?
It's not about making search smarter. It's about making reasoning visible.
03 / The asset that compounds
Rent it, and it improves someone else's product. Own it, and it compounds for you.
Every decision captured with its reason makes the next answer a little better. A shortlist drawn up against what a hiring manager actually said. A day plan that knows why an interview stalled. A new starter who inherits ten years of a client's history on their first morning.
That history is the one thing a competitor cannot buy, and a vendor should not be the one holding it. Every business already has a Company Brain. It is just scattered. The work is not building a brain. The work is consolidating, structuring and routing what is already there.
The recruitment businesses that win the next decade won't be the ones with the most AI. They'll be the ones that figured out how to capture and compound their best people's decision traces before their competitors did.
Alex Lockey, Why AI Can't Replace Your Best Recruiter (Yet)
Volume alone is not the point. a16z argued years ago that raw data makes a weak moat, because each extra record adds less. The defensible part is the decision context: why things were done, for your clients, in your markets.
An owned, documented record of how the agency decides also travels. It stays when people leave, and it is something a buyer can examine if the agency is ever sold.
04 / Context, not compute
The model changes every six months. The substrate does not.
A separation of context from compute. You can swap Claude for the next thing without losing your wiki. You cannot swap your wiki without losing your business memory.
The model is the engine. The substrate is the airframe.
Alex Lockey, Company Brain: Four Builders, One Architecture
The best model today won't be the best next year, and prices keep falling. If the layer is yours, you move to the better or cheaper model when it arrives. If it belongs to a vendor, they decide when you move, and whether you see the saving.
This is why the agency's brain connects to AI tools through an open standard, the Model Context Protocol, rather than living inside any one of them. Anthropic released it as an open standard, and it is now run by a Linux Foundation body with OpenAI among its founders, so Claude and ChatGPT both connect the same way. On business plans and the API, neither Anthropic nor OpenAI trains on your data by default. If your operational memory lives inside someone else's product, you are renting your business memory from someone whose interests are not yours.
Almost every "AI agent failed in production" story traces back to the same root cause: the agent had no substrate to operate on. The model was fine. The context was missing.
05 / Renting versus owning
Renting from your vendors feels safe. It's the expensive option.
Renting costs rise with every seat and add-on, whatever they return, and when you leave, the work leaves with the vendor. Owning costs more at the start, then grows only as you add tools that have proved their worth, and what you build stays with the agency.
Usage-based AI is a variable cost: you pay for what you use. Seat licences are fixed costs, often for features nobody touches. Vendors know the seat model is under strain: Bain found that most are adding AI usage charges on top of seats rather than replacing them. And building forces a useful discipline. The real cost is clarity. Building forces you to understand your workflows precisely.
Build first, subscribe later. You own your tools where you can. Subscriptions are rented leverage.
Alex Lockey, Mission
None of this means throwing out the CRM. Keep it if it does its job. The point is that the layer which learns from your agency should belong to your agency.
Renting from vendors
- Your data is split across each product, in each vendor's format
- What gets built is whatever is on the vendor's roadmap
- Good looks like the vendor's idea of recruitment
- The model is the vendor's choice, changed on their timetable
- You pay licences per seat, per product
- If you leave, the history and the tools stay behind
Owning the layer
- One clean record, held for your agency
- What your agency needs next, when it needs it
- Good looks like your top billers' standard
- Swappable when a better or cheaper model arrives
- Usage reported against the value each tool creates
- Your data and decision history go with you
06 / The stack shrinks
Every tool you build is one you stop renting.
Once the brain exists, each new tool starts from the same clean record and the same history of decisions. The connections are made, the data is known, and the reasons are there to draw on. So the second tool costs less to commission than the first, and the tenth costs less again.
That changes what is worth building. A product priced per seat for a job your team does twice a week can be replaced by a tool built around how your agency actually does that job, owned outright and paid for by use. Each one that proves itself is a subscription you can cancel, and the saving pays for the next build.
- 01The brain makes the next tool cheaper to build
- 02The tool takes over a job a rented product was doing
- 03It is measured against the baseline, and kept only if it pays
- 04The subscription is cancelled, and the saving funds the next build
My LLM bills are higher this month. But they're variable. I pay for what I use, and I'm constantly optimising. The SaaS subscriptions were fixed costs for features I barely touched.
Alex Lockey, The Infrastructure Flip
Over a few years the stack gets smaller and more of it is yours: the CRM, if it still earns its place, the brain, and tools the agency owns. The models stay rented on purpose, because they are the part that keeps getting better and cheaper on its own. The aim is less reliance on point-solution vendors, not none.
07 / The direction of travel
The direction is real. The deadline is not yours.
Nobody can tell you the year an agency runs differently. The mechanisms are clearer than the dates: models keep getting cheaper and better, software keeps getting cheaper to build, and the scarce thing becomes the context a model can work from. Take the mechanisms, which are free and mostly right, and set your own dates from your own position.
Current state
- AI features inside each vendor, priced per seat
- A stack of point solutions that only grows
- Data copied into every tool that asks for it
- The why lives in heads, inboxes and personal AI chats
- Spend nobody can tie to a return
- Knowledge leaves with the people who hold it
Future state
- One layer the agency owns, with tools on top
- A smaller stack, with fewer vendors each year
- Data served from one place, on request
- Decisions saved with their reasons, as the work happens
- Each tool measured against a baseline, kept only if it pays
- Knowledge stays, and compounds
None of it means handing the agency to a machine. Drafts only, at the start: the system never sends, and a human always approves. Trust is earned per task, not granted per system. The Company Brain is autopilot. Not autonomy.
How we get an agency from one column to the other, and in what order, is on how we work.
Sources and further reading
Where this comes from
Passages on this page are taken from Alex Lockey's published writing. The rest is linked below.
- Why AI Can't Replace Your Best Recruiter (Yet)
- Company Brain: Four Builders, One Architecture
- The Infrastructure Flip: From SaaS Subscriptions to Personal Software
- I Graded the Futurists
- Intelligence Withdrawal and the Leanpreneur's Playbook
- AI's Trillion-Dollar Opportunity: Context Graphs
- Long Live Systems of Record
- Big Ideas 2026: Systems of record lose ground
- Is Software Losing Its Head?
- The Empty Promise of Data Moats
- Per-Seat Software Pricing Isn't Dead, but New Models Are Gaining Steam
- Donating the Model Context Protocol
- Is my data used for model training?
- Business data privacy, security and compliance
- AI at Work Is Here. Now Comes the Hard Part
- Guidance on AI and data protection
- Requests for Startups: Company Brain
- Model Context Protocol