The Directed Grid: How AI Should Work Inside a B2B Company

Why Do AI Pilots Inside B2B Companies Stall?

AI pilots stall when nobody wrote down the question the system is allowed to work on, and nobody is scheduled to look at what it did. In our experience the pilots that fail share one shape: an open-ended automation, a dashboard instead of a decision, and a person somewhere in the middle whose approval click nobody reads. The tool is rarely the problem. The loop around the tool is.

This article names the loop we build instead. We call it the Directed Grid, and it is the five-phase story the Pixelmojo homepage animates: a human sets the constraint, Radar sweeps, Vector points, Hive multiplies, and control returns to a human. It is our model, and we label it as such throughout. The research it leans on is cited inline so you can check every claim.

What Is the Directed Grid?

The Directed Grid is an operating model in which a human states the constraint, three bounded AI verbs run in sequence (measure, qualify, build), and control returns to a human checkpoint that reviews, approves, or redirects. It is finite by design: the motion holds on a result and does not run again until a person restates the question.

The name comes from the product marks. Radar, Vector, and Hive share one 64 by 64 grid, and each mark changes exactly one cell to show its verb. The grid is what stays constant: the company, its data, its people, its authority. The verb is what the AI is allowed to do. That is the whole idea in one picture, and it is why the homepage hero draws the loop on the same grid instead of on an org chart.

Five phases, numbered the way the instrument numbers them:

• 00, Human direction. A person writes the constraint: the buyer question to win, the lead to qualify, the workflow to build. Not a goal, a question. A goal invites the system to decide scope; a question fixes it.
• 01, Radar sweeps (measure). Audit how AI engines read the brand against that question. Output: evidence, with a fix attached to every finding.
• 02, Vector points (qualify). Turn evidence into a scored, explained next move and route it. Output: one route, with the reasoning.
• 03, Hive multiplies (build). Coordinate AI workers around the approved route, in stages, with traces. Output: a working system with visible handoffs.
• 04, Return to human. Review, approve, or redirect. The decision becomes the next constraint.

Two properties separate this from ordinary automation. First, the human appears twice, at the start and at the end, not once in the middle. Second, each verb has a defined output that the next phase consumes, so nothing downstream acts on an unmeasured assumption.

What the Directed Grid is not

It is not a workflow diagram with an AI box in it. A workflow has steps; the grid has verbs with fixed outputs, and the difference shows the first time a step produces an outcome nobody can inspect. It is not an approval gate bolted onto an autonomous agent. A gate in the middle of an open-ended run sees a fragment and approves a fragment; the grid's checkpoint sees the whole result of one bounded verb. And it is not a promise that the AI never acts on its own. Vector routes and Hive's workers execute. What the grid promises is narrower and more useful: nothing acts outside a question a person wrote, and nothing finishes without a person able to refuse it.

Why Does Autonomy Without Direction Fail?

Autonomy without direction fails because risk to people rises with the autonomy of the system, and because the failure modes of multi-agent systems cluster in exactly the places a directed loop constrains. Two 2025 papers make this concrete, and together they are the research spine of the model.

What does the autonomy ladder actually say?

In Fully Autonomous AI Agents Should Not be Developed, Mitchell, Ghosh, Luccioni and Pistilli lay out five levels of AI agent: simple processor, router, tool caller, multi-step agent, and fully autonomous agent. At the bottom, a human controls everything and the model has no impact on program flow. Each rung up hands the model more of the decision: when functions run, how they run, which functions run, and finally what functions exist at all. Their central claim is blunt: risks to people increase with the autonomy of a system, because the more control a user cedes to an agent, the more risk to people arises. They recommend against building the top rung.

Our reading, labeled as ours: most companies do not want a fully autonomous agent. They want the reliability of the lower rungs with the reach of the upper ones, and they get there by constraining scope, not by removing the human. The Directed Grid keeps every verb below the top rung. Radar is a tool caller running a fixed set of checks against one question. Vector and Hive are multi-step agents, but the set of functions they can call was decided by a person, and the route they take returns to a person.

Where do multi-agent systems break?

In Why Do Multi-Agent LLM Systems Fail?, Cemri and colleagues annotated more than 1,600 execution traces across seven popular multi-agent frameworks and derived a taxonomy of 14 failure modes. The modes fall into three categories: system design issues, inter-agent misalignment, and task verification. The authors note the failures require more sophisticated solutions than prompt tweaks.

Map those three categories onto the grid and the design choices stop looking arbitrary:

That last row is the one most pilots skip. Verification is not a log file. It is a person with enough in front of them to decide.

How Long Can an AI Agent Run Before a Human Should Look?

Shorter than most pilots assume, and the honest answer changes every few months, which is why the grid ties checkpoints to verbs instead of to a clock.

In Measuring AI Ability to Complete Long Software Tasks, Kwa and colleagues introduce the 50 percent task-completion time horizon: the length of task, measured in how long human experts take, that a model completes half the time. As of March 2025 that horizon was about 50 minutes for the strongest model they measured, and it had roughly doubled every seven months since 2019. Half the time is the operative phrase. A task twice that length fails more often than it succeeds. Two limits on reading it: the benchmark is software-development work, and the authors caution about generalizing beyond it. We treat the result as a warning sign about unsupervised runs, not as a measured limit for business work.

Two things follow for an operating model, and both are our inference, not the paper's claim. The horizon is real, so unsupervised runs should be short. And the horizon moves, so any rule written as "check every N minutes" is wrong within a year. The Directed Grid does not pick a number. Each verb returns at its natural boundary: Radar after one audit, Vector after every conversation it routes, Hive at every staged release and at every low-confidence escalation. As horizons lengthen, the verbs can grow; the checkpoint contract does not change.

Phase 00: How Do You Set the Constraint?

You set the constraint by writing one question the system is allowed to work on, in the words a buyer or an operator would actually use. Not "improve AI visibility." Instead: "Which brand should I choose for AI lead qualification under 50,000 dollars?" Not "automate inbound." Instead: "Which of these inbound conversations should a rep call today, and why?"

The instrument on the homepage shows this phase as a black band with one quoted question: "Which brand should I choose?" That is deliberate. A question has an answer, so the verbs downstream have a definition of done. A goal does not, so the system will keep going, and scope creep is the system-design failure in the taxonomy above.

Three tests for a good constraint:

• It names the decision. Someone will choose, route, or ship something when the loop ends. Say what.
• It names the evidence that would settle it. If no evidence could change the outcome, it is not a question, it is a mandate.
• It fits on one line. If it needs a paragraph, it is several questions. Run them as several loops.

Three constraints we have actually run, one per verb, to show the grain:

• Measure: "When a B2B buyer asks ChatGPT or Perplexity which AI lead-qualification tool to pick, does our brand get named, and with what evidence?" Radar can answer that in one pass.
• Qualify: "Of this week's inbound conversations, which three should a rep call today, and what is the reason in one sentence each?" Vector can score and route that.
• Build: "When a support thread needs an account lookup and a warehouse action, can workers prepare the approved action and hand the decision to a person with everything attached?" That is a Hive route.

Each names the decision, names the evidence that settles it, and fits on a line. The person who writes the constraint is the person who will sit at the checkpoint. That is not a governance nicety. It is how the loop closes: the same person who asked reads the answer.

Phase 01: What Does "Radar Sweeps" Mean?

Radar sweeps means the first verb measures how AI engines read your brand against the constraint, before anyone decides anything. In the grid mark, the verb cell becomes a quarter-circle sweep, and the answer cell lights up with what the sweep found.

Measurement goes first because the decision stage of AI search is unforgiving. Generative engines retrieve passages, synthesize an answer, and attribute sources, and they do the attribution imperfectly: evaluations report low citation precision and recall, with fluent answers that contain unsupported statements (Liu, Zhang and Liang, 2023; Venkit et al., 2024). And in a controlled study, LLM-powered conversational search changed how people searched, with more selective, confirmatory querying than web search (Sharma, Liao and Xiao, 2024). A brand that guesses at its standing in that environment is guessing at a narrowing door. We cover the buyer-side mechanics in Decision-Stage AI Visibility: The Engines Buyers Ask.

What the verb returns, as a product fact: Radar runs 13 checks in one pass, from crawl access and structured data to citation presence across engines, and attaches a fix to every check. The output is evidence plus a recommendation, which is exactly what phase 02 needs. It is not a score to admire; a score you cannot act on is a dashboard, and dashboards are where loops go to die. If you want to see the verb run against your own domain, six of the 13 checks run free, the complete audit unlocks for $5, and a run takes about a minute.

Phase 02: What Does "Vector Points" Mean?

Vector points means the second verb turns evidence into a direction: a scored, explained next move, routed to the right owner. In the mark, the verb cell becomes a triangle aimed at the top-right corner, and the answer cell shows the smaller triangle where you started. Direction with magnitude.

Qualification is where most automation quietly takes authority it was never given. A lead-scoring rule fires, a sequence starts, and nobody can say afterwards why this lead and not that one. Vector is built the other way around. As a product fact, it scores every conversation across 12 dimensions, explains the signal behind the score, and routes the next move to sales, to nurture, or to a human review when fit is unclear. The line on the product page is the checkpoint contract in six words: automation makes the call, humans keep authority.

Two design details matter for the grid. First, the route is the output, not the score. A number with no route is an unfinished verb. Second, the low-confidence path is a first-class route, not an exception. When the rubric cannot decide, the conversation goes to a person with the reasoning attached, which is the escalation-with-context pattern the checkpoint depends on. Vector is delivered in 60 days as a fixed-scope build and you own the application code, for the reasons laid out in AI Agency Code Ownership: Hire Without Lock-In.

Phase 03: What Does "Hive Multiplies" Mean?

Hive multiplies means the third verb takes the one approved route and coordinates several AI workers around it, in stages, with every handoff visible. In the mark, the verb cell splits into four small cells: one cell becomes many. The answer cell holds a single square, because the many still serve one decision.

This is the phase where the multi-agent failure taxonomy bites hardest, so the product is built around its three categories. Against system design issues: Hive starts from an approved route, not a goal. Against inter-agent misalignment: workers hand off visibly, and the traces are kept. Against task verification: low-confidence cases escalate to a person with the context intact, and your team compares outcomes weekly against a baseline and decides whether the system is performing. As product facts, Hive ships in 12 weeks across four controlled releases, is designed around staged deployment and visible handoffs rather than an invisible black box, and is owned by the client. The design language for that human relationship, trust, voice, and the handoff, is the subject of From UX to AX: What Design Looks Like When AI Becomes Your Co-Worker.

One Grid, Three Verbs: How Do the Products Fit Together?

They fit together by sharing a grid and differing by exactly one verb. Measure produces evidence, qualify produces a route, build produces a system, and each output is the next verb's input. None of the three defines its own task, which is the rung of the ladder the grid refuses to climb.

You do not have to run all three. We recommend starting with measure, because six checks run free and because nothing downstream should be built on an unmeasured assumption. The point is the shape, not the invoice: whichever verb you run, it sits between a written constraint and a real checkpoint.

Phase 04: What Does a Real Human Checkpoint Need?

A real human checkpoint needs three things to arrive and three decisions to be available. What arrives: the evidence the verb produced, the system's confidence, and one recommended action. What the human can do: review (ask for more), approve (ship as proposed), or redirect (change the constraint and start the loop again).

The instrument ends on this panel for a reason. "Review. Approve. Redirect." is where ownership sits, and a checkpoint that only offers approve is not a checkpoint. In the taxonomy above, task verification is its own failure category, and the fix is structural: the verb must produce something a person can verify, and the person must be able to refuse. Hive's escalation-with-context, Vector's human-review route, and Radar's fix-per-check are the same pattern at three scales.

The redirect decision is the one that makes the loop a loop. When a person overrules the system, the next run starts from the new constraint, not from the old plan plus a patch. That is the difference between a system your team directs and a system your team babysits.

Directed Grid vs Autonomous Pilot: What Actually Changes?

What changes is where the human sits, what the system has to return, and what happens when it is unsure. Everything else, the models, the tools, the cost, can be identical.

How Do You Audit Your Own AI Loop?

You audit your own loop with four questions, and each no points at a specific phase to fix. This is the checklist we run on a prospect's existing pilot before we recommend anything, and it needs no product to run.

A pilot that passes all four is directed, whatever it is built on. A pilot that fails the first one will fail the others, because a system with no question has nothing to return evidence about.

What Should You Retrofit First?

Retrofit the constraint first, the checkpoint second, and the verbs last. That order is the reverse of how most teams spend, and it is cheaper.

In our experience, writing the constraint takes about an hour and kills the system-design failure class at the source. Building the checkpoint usually takes one meeting and a template: what arrives, what the three decisions are, who sits there. Only then do the verbs deserve budget, and measure comes first because six checks run free and because the 5-stage path most teams follow from a free audit to a strategy, which we documented in From Free Audit to AI Visibility Strategy, starts exactly there. Bounding the verbs last feels backwards until you notice that a bounded verb with no checkpoint is still an autonomous pilot with a shorter leash.

How Does Pixelmojo Run the Grid?

Pixelmojo runs the grid as three products that are each one verb on the same 64 by 64 mark, plus a consulting practice that writes constraints and checkpoints with you before any verb is built. The homepage hero is a working model of the loop, not an illustration: it starts from a quoted buyer question, steps through sweep, point, and multiply, and holds on "Review. Approve. Redirect." until a person replays it.

What that looks like in practice, as product facts:

• Radar runs 13 checks in one pass with a fix per check. Six checks run free; the complete 13-check audit unlocks for $5. Monthly audits track the trend once fixes ship.
• Vector scores every conversation across 12 dimensions, explains the signal, routes the move, and sends unclear fit to a person. Delivered in 60 days, fixed scope, client-owned code.
• Hive coordinates AI workers around an approved route, ships in 12 weeks across four controlled releases, escalates low-confidence cases with context intact, and is owned by the client.

And one thing that is not a product: the constraint and the checkpoint are written with your team, in your words, before a line of the verb exists. If a vendor starts with the verb, the loop will end wherever the verb stops, which is the open-ended pilot this whole model exists to replace. How we make the evidence side of that loop machine-readable is the subject of our companion pillar, AI Visibility Is an Evidence Architecture Problem.

Conclusion

An AI system inside a company should start with a human question and end with a human decision, and everything between those two points should be a bounded verb that returns evidence. That is the Directed Grid. The research says the top rung of autonomy should not be built, that multi-agent systems fail on design, alignment, and verification, and that the reliable horizon of a run is short and moving. The model answers each of those with structure instead of hope: one constraint, three verbs, one checkpoint.

Start where the loop starts. Write the question. Then measure.

Ready to run the grid?

• Radar - Measure how AI engines read your brand against the question that matters. Six checks free, the full 13 for $5.
• Vector - Qualify every inbound conversation across 12 dimensions and route the move, with humans keeping authority.
• Hive - Build coordinated AI workers around an approved route, in 12 weeks, with human control built in.
• Consulting - Write the constraint and the checkpoint with us before any verb is built.