Plan
of
Attack

You have no company yet, but you tell me their leadership understands the problem. How encouraging. That saves us the expense of educating them.

I have been reading The Dionysus Program, a manual for organizations trying to absorb new knowledge without destroying their ability to work together. Its instructions for repairing them make an excellent inventory of opportunities for you.

Start with the regional home-goods retailer you mentioned. Its buying manager receives a monthly stock report from an inventory consultancy. The consultancy has tested daily warnings about shortages across her stores and shown they arrive in time to prevent them. She has seen the trial. She still cannot buy the service.

Find someone who can substantiate her account of the trial and explain why it stayed a trial. Perhaps the senior reviewer derives standing from signing every report; the partner’s relationship rests on being the person who knows. Daily warnings make both arrangements difficult to preserve. This is epistemic melt: the knowledge has moved; the careers and relationships built on it cannot move as quickly.

An honest executive may spend a year helping decent people find their place in the new work. Respect the difficulty. It is paying your expenses.

But their successful trial is not yet your business. Reproduce the result with data the customer can give you, expertise you can afford, and a buyer willing to act without the incumbent’s name behind the advice. Count the unusual cases, missed warnings, and cost of putting errors right. A cheap demonstration is often a bill with several pages missing.

Then ask the buyer to pay for a limited trial and renew after seeing the result. If the improvement depends on the incumbent’s exclusive data, trusted specialists, or another service you cannot supply, its internal quarrel has furnished you with an interesting evening.

Suppose the result does travel. Arrange the work around it.

AI agents—software permitted to take actions—can check inventory and issue warnings wherever your trials establish that they do so reliably. Give them usable instructions: which evidence justifies a warning, what errors cost, and when to stop. People choose what the service promises, investigate failures the rules cannot explain, and take responsibility for the consequences. Making the software await your blessing on every ordinary warning would recreate the arrangement you came here to attack.

Let corrected mistakes change the instructions and tests. Give the investigator the next unsolved problem, rather than a permanent desk between you and the software. None of the old firm’s careers exists in your company yet. Be economical about creating them.

Every person you place permanently between a decision and its execution is a piece of the incumbent you have begun building inside your own company.

At renewal, make the customer ask why the service her supplier validated is available from you. Expect a discount, a new tier, or an independent operation. If the rival routinely matches your result at a comparable cost, its partners’ unhappiness is no longer your business model.

Let the incumbent keep the distinction of having seen the future first. Send the renewal proposal.

Your inventory warnings have earned the retailer’s trust. Now the buying manager wants you to arrange the earlier deliveries your warnings call for. You propose hiring someone to interpret your wishes for the software and carriers. Keep the job description. We shall return to it.

Ask the buying manager for the histories of requests to change a delivery with her logistics provider. Locate the wait after the facts are known but before anyone may commit. A missing part or a genuinely novel decision is a different problem. Your enthusiasm cannot make either into a clerical error.

One store needs Thursday’s delivery on Tuesday. The goods are ready at the retailer’s warehouse; transport will cost another $200; the buyer accepts the surcharge. The provider’s sales and dispatch teams still need their chief executive to approve the combination. Everyone has prepared the decision. Only the person who sets the company’s direction may finish interpreting it. An Auteur Bottleneck is particularly well defended when the person at its center is excellent. A gifted one can keep an entire market waiting while appearing to save it every afternoon.

First secure the retailer’s authority to arrange shipments and contracts with carriers whose capacity you can reserve. Your software can then settle that recurring case under explicit instructions: reserve the goods and Tuesday capacity, preserve the promised scope and minimum margin, and confirm only if no existing commitment is displaced. If Tuesday requires taking another store’s reserved stock, the surcharge does not make the answer yes. Offer Thursday, or send a genuinely unresolved tradeoff to someone authorized to judge it.

Now consider your proposed interpreter. Every ordinary Tuesday order would still need someone to explain what you meant. That is your bottleneck with a salary.

Test those decisions against completed jobs and deliberately awkward combinations. Sell a changed delivery with a confirmed date and price while the rival’s executive is still reading the request.

There is a trap here. If a published price band repairs the rival’s problem, your opening is slight. Find out whether sales is rewarded for accepting work that dispatch is punished for accommodating, and whether anyone below the executive can settle that conflict. Customers can show you the queue; people who work there must help you establish what keeps it there. Matching you may require changing both authority and rewards. It may also happen sooner than you would enjoy.

Where parts of a job continually change one another, bring the responsible people together to settle the whole commitment. Give their judgment somewhere useful to go: a solved class of cases should become executable instructions. An exception should have a chance to graduate.

Prove delivery across repeated jobs, different conditions, and your absence. I am particularly interested in your absence. Founders conceal a great deal of defective organization by declining to sleep.

Employ your proposed interpreter’s judgment on decisions the rules cannot settle. Tuesday’s order should go through while you are both doing something else.

Watch the queue. Yours, too. Your success should give the rival’s executive more to decide and you less.

The retailer now wants to move its regular store deliveries to you, alongside the urgent ones. Before you celebrate taking that business from its logistics provider, send me the order history that made the buyer ask. I have a weakness for endings.

Here is the pattern to look for: sales software confirms Friday delivery at the quoted price. Scheduling reserves a truck for Monday. An automated support agent, instructed to preserve the original promise, restores Friday without securing a truck. The store keeps two employees late to unload. Nothing arrives.

Each agent may obey its instructions perfectly. The instructions may be impossible to honor together. The provider has crossed its Coherence Ceiling: more agents no longer get more orders delivered. This one adds an assurance and a repair bill. Establish whether such additions cost more than the useful work they contribute. General dissatisfaction with a large company affords you no special insight.

Now I must disappoint you. The rival can cut activity, restrict permissions, repair connections, or redesign the work. Turning off one agent may cure an apparently magnificent strategic weakness. Check before ordering the champagne.

Suppose instead that every repair ends at a conflict nobody can settle. Sales retains authority to promise dates that dispatch cannot supply. Support must honor them. An operations employee can show you the contradictory rules; customers can show you the resulting reversals. Management adds human review, sending every exception to the chief executive who already approves delivery changes. Its repair has joined the queue.

There is your offer: a date, scope, and price the customer can rely on, confirmed soon enough to be useful. Your advantage lasts only while the rival cannot match the whole service at a comparable cost. Its embarrassment alone pays poorly.

Take the same order through your own operation. A date cannot be offered until the required stock and transport are reserved. Support can change the commitment only by changing those reservations too. If the customer wants something those rules cannot resolve, someone must judge the consequences and take responsibility. Three fluent answers do not constitute a delivery.

When you change the rules, withdraw the obsolete permissions everywhere they could still make a promise. Try to provoke the old behavior. A correction is unfinished while some agent can sell the previous promise.

Compare repeated orders through delivery: elapsed time, reversals, repair cost, and the customer’s own effort. Include the whole job. Leaving its difficult half with the incumbent can make you look wonderfully well coordinated. I should prefer your first profitable illusion to belong to someone else.

Only then increase the number of agents permitted to act. When commitments collide, reduce the load and repair the cause. Another agent is an extravagant way to postpone admitting that the instructions disagree.

The customer has already paid for a promise. Offer the unusual experience of receiving it.

The retailers whose deliveries you manage have given you another opening: returns. Your first correction there worked. You are delighted that their old returns contractor needed six months to follow. I shall be more delighted when you can explain why its next change will take six months too.

Choose work in which discoveries repeatedly change the method and who may decide. In stable work, a rich incumbent may repair itself once and leave you with a charming account of its former difficulties.

The contractor held every refund for inspection at a depot. Your trial showed that photographs reliably cleared a defined class of ordinary returns before the parcel arrived. The retailers authorized you to issue those refunds within agreed limits. An experienced investigator handles ambiguous evidence and challenges the limits when results warrant it. Faster refunds won you the work; the difficult returns still have to reach an inspector. Price that service too.

The rival’s inspectors had to surrender their veto over those returns. Their standing and remaining work had to change. Follow the delay from usable evidence to refunds actually arriving sooner. An announcement does not stop the clock.

Then the completed returns reveal another mistake: some goods cost more to retrieve than they recover. You test waiving collection for those cases, counting fraud and lost resale value as well as transport savings. If it works, the service can refund promptly at a lower cost. The rival must now confront a depot whose manager is rewarded for parcel volume. Sell the saving while it settles another change of authority.

You need both corrections to reach ordinary customers, repeatedly, at a profit. A succession of demonstrations proves that you enjoy demonstrations.

From the first return, keep the results needed to test the rule that produced them. A proven correction should change the software’s instructions without employing a permanent human reviewer for every parcel.

Improving the machinery of correction takes another test. Suppose the software notices that its tests omit collection cost and proposes adding it. Try the revised test on completed returns it has not seen, then check limited live use against actual losses and savings. If later changes are judged more accurately, you have begun recursive improvement: improving the means of improvement itself. Extend its authority only where those checks hold up. Faster refunds remain part of the promise; software may not discover profitability by quietly ceasing to provide them.

Your investigator will eventually challenge a rule you wrote. Work beside them on difficult cases, respect their craft, and spend some time together that earns nobody a score. Trust needs a life outside the meeting in which you intend to spend it.

When a rule needs questioning, suspend the ordinary output demands long enough to examine it. Require evidence and responsible conduct; spare exhausted people another compulsory revelation. Retire the failed rule without humiliating its author. Look for the next inconvenient fact offered freely. That is the test of epimetabolic rate: how quickly error becomes better work without making the next truth harder to tell. A faster fix followed by silence has borrowed against your future.

Watch whether the rival starts correcting faster while keeping its critics. Your lead can shrink.

Make it pay for another transformation while you are already delivering the result. Arrive at the next error with people still willing to spoil your afternoon.

Your retailers now ask why the shipment records you hold cannot settle the carrier invoices they spend days disputing. You have found a candidate from their reconciliation service. The account of being misunderstood was moving. I hope you allowed them to finish before asking what, precisely, they were right about.

Their account is specific: the service sends every mismatch back to the retailer for investigation. Its matching rule treats the second truck on a split shipment as a duplicate charge. The candidate proposed resolving those cases from records already available. The manager who wrote the rule refused a test and removed the candidate from account reviews.

Ask for evidence of the error, the proposed remedy, and the refusal. Reconstruct the strongest reason to say no. Perhaps the records omit an obligation. Perhaps the method fails in ordinary use. I am fond enough of your ambition to insist that it stand on something firmer than another person’s grievance.

Before you offer a title, offer a bounded trial. Give the candidate authority to demonstrate the method and responsibility for its consequences. A buyer must be willing to entrust you with the necessary records and pay for the corrected result. Price the checking, failures, onboarding, and support. Then ask for another paid batch. A cheaper internal process does not necessarily support a business outside the firm that supplied its customers and infrastructure.

Observe what happens when evidence damages the candidate’s explanation. Someone who can correct everyone except themselves brings an interesting succession problem.

If the trial survives, sell a precise offer: carrier invoices reconciled by a stated deadline, including split shipments, with doubtful cases investigated before approval. Win the next account with completed work and an honest account of what it cost the buyer to use it. Being vindicated is a private pleasure; customers are purchasing something else.

Now decide what the recruit does after the method is understood. Put reliable repetition into instructions the software can execute. Employ the discoverer to investigate what still goes wrong, judge doubtful cases, and develop the next correction. A permanent right to be correct would be a peculiar signing bonus.

The previous employer may offer genuine influence. If the rejected method can now receive a fair test and its result can change the work, your recruit has a serious reason to stay. If critics are still punished, its negative selection is your recruiting service: the firm drives out people who could repair it and teaches everyone remaining which truths are expensive. As you win business, it may lose the people best equipped to explain why.

Eventually your recruit will question your favorite assumption. You will be busy. Their timing will be poor. You will have a remarkably persuasive explanation for why this criticism is different.

Give the question a fair test anyway. Let a sound result change the work, and let a failed experiment leave the person with a future. You hired someone who could tell an unwelcome truth. I should hate to discover that your offer applied only to truths about somebody else.

You now move goods and reconcile the bills for regional retailers at a profit. Independent shops want the same frequent deliveries. This is when I begin to worry about your explanations.

The large distributor serving those shops has published a report concluding that this market does not exist. Keep a copy. It may prove a useful record of what its machinery allowed it to know.

Its premise is that small buyers have no use for frequent deliveries. It excludes them from the offer. Automated sales systems classify their requests as poor fit. Forecasts describe only customers who passed the screen. The reports are accurate about the population the rule permits them to observe. Management has arranged its evidence beautifully.

Start with the shops that approached you. Ask what they requested and what happened next. Look for written refusals of twice-weekly deliveries and someone who can establish whether trials are forbidden below the existing account minimum. Where small accounts cannot qualify for a test that could show why the minimum is mistaken, the rule has acquired custody of the evidence against it.

Now establish what the distributor has prevented itself from learning. Your margins so far came from services; buying stock means financing it and bearing the loss when it does not sell. Pool firm shop orders to negotiate terms with independent suppliers, then count the cash tied up, storage, returns, and unsold goods before admiring the margin. Serve a defined group at the requested frequency and obtain repeat purchases that cover the full cost, including the awkward stops. If that contradicts its premise, you have found coherent wrongness: a whole firm acting consistently from a false account of the world. Improve your offer while its reports continue finding no demand.

You need the false assumption, a result that contradicts it, and the mechanism that prevents correction. Mere refusal is insufficient. The firm may understand the demand and decline it for sound reasons of cost, obligation, or better opportunity elsewhere. Your margin does not prove its decision foolish. Accepting the evidence but struggling to reorganize is another weakness. Here the firm keeps the question from receiving a valid answer.

Your protection ends when contrary evidence can reach a fair test and its result can change the rule. You may still have a lead; you can no longer rely on the rival’s ignorance to preserve it.

Which brings me to your profitable segment. You will want your software to find more customers like those customers. Sensible. Reward it for rejecting everyone else, and each successful sale will confirm the wisdom of its selection. A self-improving operation can become very accomplished at polishing its blind spot.

When it declines a request, keep the request and the reason. Suppose someone asks for weekly service and your rules reject it as unprofitable. Give the person investigating such refusals access to the omitted cases and a modest budget to test that assumption. The investigation should not await the afternoon when curiosity becomes your highest priority.

If repeat purchases overturn the assumption, change the rule and the tests that reward the software for enforcing it. Check the revised tests against actual customer results, including the people previously excluded. A test declaring itself improved is charming correspondence between machines. Someone remains responsible for what the company promises and for consequences those machines cannot judge.

Your first profitable segment will furnish a persuasive story about what the company is. Your software will tell it back to you with increasing fluency. Arrange now for evidence from outside that story to interrupt.

I should like these letters to remain useful to you. It would be a pity if their next reader found them useful chiefly as a description of your company.