Journal
The Brainis blog
Understand, decide, do, check, learn. Notes from the people building Brainis, and running our own company on it.
48 articles
Governed autonomy
What should an AI circuit breaker do when it trips?
An AI circuit breaker should limit further action in stages. Learn what it watches, when to throttle, pause, or stop, and what to review before work resumes.
Company intelligence
How to choose company health metrics that guide action
Choose company health metrics by starting with the decisions you face, linking each measure to an action, and noting what it cannot show on its own.
Deciding with models
Work without deposit: should you accept the client job?
A work without deposit decision weighs cash at risk, the client's commitment, and options if payment slips. Compare terms, set a limit, and record your call.
Strategy to execution
Product line investment: which line gets the next dollar?
Use product line investment to compare strategic fit, evidence, and downside, then set a gate for where the next dollar goes—and when to stop.
Running the company
Managing too many projects: rebalance one person's work
Too many projects do not prove one person is overloaded. Map ownership, dependencies, and priorities, then decide what to move or pause and review the result.
Category
How to track AI margin by customer and feature
See how AI margin by customer joins AI usage to revenue by feature, preserves cost lineage, and helps you review negative margins and proposed fixes.
Verified AI work
An AI analysis checklist: what to inspect before acting
Use an AI analysis checklist to inspect its inputs, method, uncertainty, and limits before deciding whether the recommendation is ready to act on.
Starting up
When to add a paid revision cap to a service offer
Learn when a paid revision cap fits, how to define an included round, price extra work, write the boundary into your offer, and review whether it worked.
Governed autonomy
What autonomy grant changes mean for a running mission
Autonomy grant changes don't reset running missions: they finish under their starting grant; new acts need the current level. See what carries over.
Company intelligence
Company memory onboarding: what new hires need first
Company memory onboarding should give new hires the picture, the reasons behind decisions, and where each fact came from—so they know what to trust and update.
Deciding with models
How to make a conflicting evidence decision before acting
A conflicting evidence decision starts by naming the disagreement, checking source and freshness, and setting a rule for what is strong enough to act on.
Strategy to execution
Mission blocker response: what should happen next?
A mission blocker response names the blocker, assigns an owner and an SLA, and sets an escalation path. Learn what to record so a decision can let work resume.
Starting up
Why hiring more engineers won't speed up software delivery
Why hiring more engineers fails to speed up releases, how unseen code review queues and dependencies block work, and how to fix your binding constraint.
Verified AI work
AI verification failure: What to do after Verify rejects it
When an AI verification failure occurs, use the rejection findings to locate the failed check, revise the deliverable, rerun relevant checks.
Governed autonomy
Managing AI agent fleets with scoped permissions
Managing AI agent fleets requires scoping tool access and context per mission. Learn how to govern model permissions without granting full system access.
Starting up
How to raise service prices without guessing wrong
To raise service prices, compare delivery effort, demand, and costs; choose one service for a test, then track the response before extending the change.
Company intelligence
How company state privacy protects your sensitive data
Learn how company state privacy separates structure from raw text to give you AI foresight without exposing confidential client or internal data.
Deciding with models
How to simulate business decisions before risking capital
Learn how to simulate business decisions before committing real capital: set baseline state, test alternative strategic branches, and evaluate tradeoffs.
Strategy to execution
How to set abort conditions for projects before launch
Learn how setting clear abort conditions for projects before launch prevents emotional drift, controls spend, and stops failed initiatives without blame.
Running the company
Pipeline capacity planning: predicting when to hire
Learn pipeline capacity planning to predict hiring needs before signing big deals. Map delivery workloads, spot bottlenecks, and time recruitment accurately.
Running the company
Acquisition channel profitability: finding true net losses
Evaluating acquisition channel profitability requires joining customer revenue with model usage and delivery costs to reveal negative-margin channels quickly.
Category
Why an AI operating system beats fragmented AI tools
Point AI tools create disconnected outputs. An AI operating system connects state, decisions, missions, and verification into one continuous feedback loop.
Verified AI work
How to define done for AI work that actually holds up
When you define done for AI work using explicit acceptance criteria and automated verification, incomplete deliverables stop slipping into production unnoticed.
Verified AI work
What our Verify layer rejected this month (dogfood report)
A report on the shape of what independent verification caught in our own company this month — the categories, not counts. We do not publish numbers we cannot yet source from live records.
Running the company
Referrals-to-zero: the anti-agency metric
Every time we hand a customer to a human service because our software cannot do the job, we record it. Driving that count toward zero is the only metric that keeps a software company from becoming an agency.
Starting up
Pricing your first pilot
Free pilots teach you nothing about willingness to pay and cost you the strongest signal available. A small paid pilot is a different conversation with the same people.
Starting up
Starting an AI B2B SaaS in 2026: the 30-day design-partner path
Thirty days is enough to find out whether the thing you want to build is a business. It is not enough to build it, and confusing the two is the most common way the first month gets wasted.
Starting up
Ten design partners before a landing page
A landing page tests whether your copy is good. Ten design partners test whether the problem is real. Doing them in the wrong order is how a year gets spent optimising a headline for a product nobody needs.
Running the company
The binding constraint: finding the one thing to fix
At any moment one thing limits your company's output. Improving anything else produces activity and no result — which is why busy quarters and flat quarters correlate.
Strategy to execution
The eight artifacts every launch actually needs
Not a business plan. Eight specific artifacts, each built from a named framework, each answering a question you will otherwise answer by accident six months in.
Governed autonomy
What 'earned autonomy' means in practice
Autonomy that is granted because a vendor said it was safe is a guess. Autonomy that is granted because a domain's own record justifies it is a decision — and the record is the product.
Company intelligence
The morning brief that replaced our status meetings
We stopped running status meetings and replaced them with one composed brief: five sections, no more than seven items, every number carrying its evidence class, every line deep-linking to the thing it describes.
Deciding with models
Why we write predictions down before we act
Memory rewrites forecasts to match outcomes. The only defence is a record made before the outcome exists — sealed, so it cannot be edited into agreement afterwards.
Running the company
Coexist, orchestrate, consolidate: adopting AI without rip-and-replace
Nobody switches their whole company at once, and any vendor whose plan requires it has not run one. Three phases, in order, each of which is a stable place to stop.
Category
Why AI agencies won't scale
An AI services business grows by hiring. That is not a criticism of the people — it is arithmetic, and it decides what the business can become.
Strategy to execution
From business plan to mission graph in one afternoon
A plan is prose. A mission graph is prose with dependencies, owners, budgets, acceptance criteria, and abort conditions attached. The conversion is where most of the thinking actually happens.
Verified AI work
Evidence graphs: proving AI work happened
A deliverable is a claim. The evidence graph is what backs it: which agent produced it, under whose authority, against which criteria, checked by what, at which revision.
Deciding with models
Ranges beat certainty: calibrated decisions for founders
A point estimate hides the thing you most need to see: how wrong it could be. Ranges force the conversation to the width, which is where the decision usually lives.
Governed autonomy
L0–L7: how much should your AI do?
Eight levels, from observe-only to operating a scope end to end. The useful question is not which level is best but which level each domain of your company has earned.
Company intelligence
Fact, assumption, prediction: teaching your company epistemics
Most bad business decisions are not reasoning failures. They are classification failures — an assumption that got treated as a fact. Three labels, applied consistently, change how a company argues.
Strategy to execution
Your strategy deck is already stale
A deck is a snapshot of a belief, rendered as slides, disconnected from the state it described. By the time it is circulated the state has moved and the deck cannot know.
Verified AI work
The agent that does the work shouldn't grade it
Self-assessment by a language model is not a check. It is the same reasoning that produced the work, asked a second time in a friendlier tone.
Deciding with models
What a Business World Model actually predicts
Not the future in general. A world model for business predicts the consequences of specific decisions, in ranges, against a state it can see — and seals each prediction when it is made, so it can be held against what actually happened.
Governed autonomy
Autonomy is a contract, not a switch
The question is never whether the AI can act. It is which actions, in which domain, up to what value, at what risk, reversible or not, within what window, on whose authority.
Category
Software all the way down
The test is not whether AI is involved. It is whether, at every layer of delivery, the thing doing the work is software operating under rules you can inspect.
Company intelligence
What is company state (and why dashboards aren't it)
A dashboard shows you numbers. Company state is a live model of what is true about your business, with a source, a freshness stamp, and a confidence level on every field. The difference decides what your AI can safely do.
Category
The company loop
Understand, decide, deliver, verify, learn. Every company runs this loop; most run it in pieces, across disconnected tools, with the learning step missing entirely.