01
Which agents, in what order
Agent projects fail on selection, not engineering. Every candidate has to clear four tests before anything gets built, and we tell you what fails them and what would have to change first.
Why us
The world is full of AI services. Three things separate us.
-
Architects, not another engineering shop
Plenty of firms will build you an agent. Far fewer can tell you which one is worth building, in what order, and where a human still has to stay in the loop. That judgment is the work, and it is what decides whether the build pays.
-
Operators, not consultants
We ran revenue teams before we advised them. Customer success, revenue operations, and go-to-market leadership inside companies scaling from twenty million to several hundred million. We have made these calls with our own numbers on the line.
-
Adoption is part of the build
An agent nobody uses is a failed agent. Change management and coaching are inside the engagement rather than sold separately, because what stalls these deployments is almost always people and process rather than pipelines.
You should be building in house. We are the accelerant, not the substitute.
Internal ownership is the goal, not the fallback. Boards and CEOs are increasingly saying the same thing: we should not have to hire an outside firm every time we want to deploy AI. They are right. Our job is to be a force multiplier on the team you already have, so you get to a foundation your own people own and run with fewer wrong turns than doing it cold.
The first agent is the expensive one because it carries the enterprise foundation: identity, permissions, connectors, governance, auditability, deployment patterns. Most vendors treat that as the cost of implementing their product. We treat it as the beginning of a capability, and it is the last time anyone pays for it. Everything runs in your cloud, you own the artifacts, and the connectors, registry, and audit ledger become the substrate your own engineers build on.
Which means the thing you are actually buying is the judgment about what to build, and the change management to make it stick, not the capacity to build it. We are architects, not another engineering shop, and in an advisory engagement we do not take the keyboard. Your engineers build. We make sure they build the right thing, in the right order, on a foundation that holds. Same scorecard. Measure us both.
The premise
Why the value is stuck between your systems.
Six lines. If you disagree with any of them, we are probably not the right firm.
- Your most expensive revenue problems live in the handoffs, not inside any one system.
- For twenty years the only way to fix that was more people.
- AI is the first thing that changes the math, and the handoffs are where it pays first.
- Which is why you are not buying AI. You are buying revenue per employee and a lower cost of growth.
- Anyone can build the agent now. Knowing which work is safe to hand off is what decides whether it pays.
- We build across the stack you already own, and you keep the foundation, so the second agent is cheaper than the first.
How to engage
Two shapes. Your team owns the outcome in both.
You can switch. If you are working with us monthly and decide to hand over a defined build, we pause the retainer and credit what is left toward the project fee.
Your team builds. We architect.
A monthly engagement for teams who have the capacity and want the architecture and sequencing judgment. We do not take the keyboard. Your engineers build, and we make sure they build the right thing, in the right order, on a foundation that holds.
What you get each month: both principals, a weekly working session with your engineers, a strategy session with your GTM and technology leaders, async access in between, and named artifacts agreed in advance.
Artifacts might include an agent portfolio map with sequencing, a reference architecture for the foundation, a use case brief for the next agent, a human-in-the-loop boundary spec, or an adoption plan for an agent already live.
We build. Your team owns it.
A defined project for teams that need it moving faster than their own capacity allows. Fixed scope. A working sandbox on your own data in two to three weeks, then the agent, then the handover: the running system, the runbook, and the architecture.
Your engineers are in it with us throughout, because the whole point is that they own and extend it after we go. No license, no seat count, nothing to renew.
What we believe
Four principles that shape every engagement.
The product is the architecture. The agent is the artifact.
Agents are table stakes. Architecture is not.
Agent building is becoming a commodity skill, and most companies will have several before long. What separates the ones that compound from the ones with a pile of pilots is whether anyone designed the architecture underneath: which agents get built, in what order, where humans stay in the loop, and whether the revenue engine reshapes around any of it.
We decide what to build before we build it.
Whether an agent creates value or quietly destroys it is rarely a technical question. Every candidate has to clear four tests before we scope it, and if it does not clear them we say so and tell you what would have to change first. Naming what we will not automate is how we earn the right to be believed about what we will.
Humans decide. Agents do the work.
Automating the decision instead of the work is the expensive mistake, and there is now research behind that claim rather than just opinion. Agents detect, generate, and draft. Humans decide. The handoff is designed on purpose, and it is what keeps a deployment from silently failing eighteen months in.
Revenue per employee is the scoreboard.
Not seats, not user counts, not AI in the workflow. Revenue and profitability per employee is the number that captures growth and cost discipline in one line, which is why it keeps showing up in how investors compare portfolio companies. If a deployment does not move it eventually, it was a science project.
How we work
Sandbox. Build. Hand over.
Fixed scope, written exit criterion, no open-ended consulting engagement. The output is a working agent in your cloud, not a slide deck.
-
Sandbox
We run the candidate against the four tests and stand up a working sandbox on your own data, redacted, in two to three weeks. Not demo data, your accounts. Fixed fee, fixed duration, and if nothing clears the tests the recommendation is do not proceed and the sprint fee is all that is owed.
-
Build
The first agent carries the foundation: identity, permissions, connectors, governance, audit. It runs in your cloud on the stack you already own. Weekly working sessions with your revenue leaders, and both principals on the engagement rather than a delivery team behind a pitch team.
-
Hand over
Your team inherits the running system, the runbook, and the architecture you paid for once. Future agents extend the substrate instead of rebuilding it. We measure against the metric we underwrote to, compared with a matched control cohort, and stay through the first ninety days.
The four tests every candidate agent has to clear
Selection is where agent projects live or die. These are the questions we answer before any code gets written.
- Is the data clean enough that a decision made on it is a real decision?
- Is the decision actually decidable, or does it depend on context that only lives in someone's head?
- Is there a defined action waiting on the other side of the decision?
- Is anyone accountable for the outcome once the action is taken?
Questions B2B revenue and technology leaders ask us.
Answered the way we would answer them in the room, which is to say specifically.
How do you determine which AI agents are actually worth building?
Every candidate has to clear four tests before anything is scoped: is the data clean enough that a decision made on it is a real decision, is the decision decidable without context that lives only in one person's head, is there a defined action waiting on the other side, and is anyone accountable for the outcome once the action is taken. Candidates are then ranked by the size of the trapped value against how many tests they clear today. AI agent projects fail on selection far more often than on engineering.
How do you evaluate if a process is ready for an AI agent?
Readiness is about legibility rather than maturity. We trace the process end to end, check whether the fields the decision depends on are populated and consistent across systems, whether the trigger is observable in data rather than in someone's judgment, and whether the downstream play already exists. Where a process fails, we say which specific gap has to close first, so the fix is a scoped piece of work rather than a transformation program.
Should we build AI agents in house instead of hiring a partner?
You should build in house, and this makes that cheaper rather than redundant. The first agent is expensive because it carries the enterprise foundation: identity, permissions, connectors, governance, and audit. We build that once, inside your cloud, and your team inherits it. Every agent after the first reuses it, including the ones your own engineers build. Same scorecard, measure us both.
What does an AI architecture advisory retainer include?
A monthly engagement with both principals. A weekly working session with your engineering team, a monthly strategy session with your GTM and technology leaders, async access in between, and named artifacts agreed in advance each month. Typical artifacts are an agent portfolio map with sequencing, a reference architecture for the foundation, a use case brief for the next agent, a human-in-the-loop boundary spec, or an adoption plan for an agent already running. We do not take the keyboard in this shape. Your engineers build, and we make sure they build the right thing, in the right order. Two-month initial term, then month to month, and the commitment credits toward a project fee if you later hand us a defined build.
Can you work alongside our internal team instead of building it for us?
Yes, and it is often the better shape. If your team has the capacity and wants the architecture and selection judgment, we work monthly: both principals in a weekly working session with your engineers, a monthly strategy session with your GTM and technology leaders, async access in between, and named artifacts each month agreed in advance. We are not hands on keyboard in that shape. Your team builds, and we help them build it right, faster. If it needs to move faster than your team’s capacity allows, we take a defined scope instead. You can switch between the two, and a monthly commitment credits toward a project fee.
Do we own the AI agent architecture once the engagement ends?
Yes. Everything runs in your cloud and you own the artifacts: the agent, the connectors, the registry, the audit ledger, and the runbook. There is no license, no seat count, and nothing to renew. Future agents extend that substrate instead of rebuilding it, which is why the second one costs a fraction of the first.
What is the human-in-the-loop handoff in B2B GTM?
It is the designed boundary between what the agent does and what a person decides. Agents detect, generate, and draft. People decide, especially on selection and on anything with a customer or revenue consequence. A field experiment with 791 professionals at Procter and Gamble, published in Organization Science in 2026, found that people working with AI picked their own best idea about 37 percent of the time against roughly 50 percent for teams without it. Output was still better overall, but the failure mode is clear: automating the decision instead of the work.
How do you ensure AI agents do not make costly autonomous decisions?
Three mechanisms rather than three adjectives. Every output is cited back to source evidence, and a separate verification step blocks any claim it cannot support. Every decision is written to an append-only audit ledger, so an action can always be traced. And the agent's authority is scoped at build time, with a named human owner on any action that touches a customer, a price, or a contract. We also commit to a false-alert ceiling, because a noisy system gets ignored, and an ignored system is a failed one.
Does this add another SaaS contract to our stack?
No. The agent runs inside your own cloud on the systems you already own. Salesforce, Gainsight, Zendesk and the rest stay your systems of record. Nothing leaves your environment except the model call, your secrets management stays yours, and your IT team referees the configuration.
What makes this different from standard B2B SaaS or RevOps consulting?
A SaaS vendor can only automate inside its own product, and the most expensive revenue friction sits in the handoffs between products. A consultancy diagnoses the friction and hands over a document, so nothing compounds and the client pays again next time. We decide which work is actually safe to hand off, build the agent across the systems you already own, and leave behind a foundation that makes every later agent cheaper, including the ones your own team builds.
How would we know it worked?
Against the metric the agent was chosen to move, and ultimately against revenue per employee. Outcomes are compared with a matched control cohort of the same segment, size band, and renewal quarter, because an annual number on its own cannot establish causality.
Who we are
Two operators who ran the seats we now build for.
The judgment works because we have made these calls from the operator side, with our own numbers on the line. Both principals are on every engagement, from the first working session through the ninety-day handover.
Beth Yehaskel
Principal
Twenty-five years running revenue teams. VP Customer Success and then VP Business Operations at Spredfast, where the business scaled from roughly $20M to $60M ARR and integrated two acquisitions. Head of Customer Success for the Americas at Dropbox. VP Customer Success at Jungle Scout, leading a churn turnaround through hypergrowth.
Revenue Architect at Winning by Design from 2020 to 2025, teaching more than 3,000 go-to-market leaders, with revenue architecture work across companies from $20M to $350M ARR. ICF-certified executive coach.
At Cognitive Edge: post-sale and retention mechanics, revenue-architecture-driven agent scoping, the four tests, and the change management that decides whether any of it gets used.
Ernesto Humpierres
Principal
Over a decade across B2B SaaS, in both operating and advisory roles. Formerly at Atlassian in revenue and go-to-market roles, then commercial leadership inside PE-backed portfolio companies, working directly with private equity operating teams on revenue performance.
At Cognitive Edge: revenue engine analysis, agent selection, deployment architecture, and the operating changes that make the agents stick.
Get your team moving faster.
Tell us roughly where your build is. Both principals read every one of these, and you will hear back from one of us, not a sequence.