You’re the owner. Someone on the team has an AI subscription. A vendor has shown you a convincing demo. The office still runs on the same inbox, spreadsheet, and person who remembers where everything goes. You can see several possible uses for AI, but choosing one feels harder than buying access to all of them.
Before recommending work, we look for the conditions that could make a useful idea stall. Use these eight patterns to describe your starting point:
- No clear starting point
- Untrained teams / no standards
- Information scattered / no source of truth
- Spreadsheets beside the system of record
- Overlapping subscriptions
- Manual reporting
- Tried before; didn’t stick
- Want guardrails / human approval
This is a recognition checklist, not a score that diagnoses your business. One consequential problem can deserve attention; several familiar patterns do not automatically justify a build. You may need team practice, a clearer process, a connection between tools, or a decision to leave something alone.
Applied Margin’s job is to understand the work and help the team use the right change reliably. We trace the task, choose a useful intervention, build when the case supports it, and prove that people can operate it. The examples below illustrate what to examine, not client histories or promised results.
1. No clear starting point
You want the office to use AI, the reports to arrive sooner, and the follow-up to stop slipping. Asked which matters most, you would select all of them. In an HVAC business, the dispatcher and estimator might each have a different answer, while the owner sees the cost of both.
Starting with a tool leaves those priorities unresolved. A team can spend its energy on the easiest demo while the expensive bottleneck continues unchanged.
What we do first: use a free introductory conversation to understand the business and choose the next question to answer. The AI Workflow Sprint fits when the team needs practice on real tasks. An optional Margin Scan fits when an operating area needs deeper investigation. A solution you have already mapped can move directly into implementation scoping.
Read When a Paid AI Assessment Is the Right Starting Point for the distinction. Paid discovery is an option, not an admission fee to every engagement.
2. Untrained teams / no standards
One person pastes notes into a model. Another writes a fresh prompt every time. A third avoids AI because nobody has explained what information is appropriate to share or how to check an answer. The agency’s weekly client update looks different depending on who drafted it.
Without a shared workflow, quality depends on individual habits. An apparently polished draft can hide missing facts, and nobody knows which version of the instructions to improve.
What we do first: work on agreed tasks in the AI Workflow Sprint, $1,500 fixed. It covers one team of up to five people, a short leadership kickoff, two hands-on working sessions, two to three reusable workflows, written instructions, an opportunity map, and one follow-up session. Session lengths and scheduling are agreed before booking. Software, integrations, and subscriptions are separate; there is no automatic credit toward a later build.
Practice follows the Operator Standard: understand the task, use the workflow, verify the result, and recover when something goes wrong. Read Buying ChatGPT for Your Team Is Not an AI Strategy.
3. Information scattered / no source of truth
The approved procedure is in a drive folder. The latest client detail is in the CRM. The exception lives with the office manager. In a business with several locations, people can answer the same question differently while each believes they have the current information.
An assistant cannot resolve conflicting sources simply by reading more files. It needs agreed materials, an owner who can correct them, and rules about who may see what. Otherwise, it can make an outdated answer easier to repeat.
What we do first: identify the authoritative material and its owner. Sometimes the two-file pattern is enough to document basic company context and customer expectations. When a queryable system is useful, Implementation starts at $3,500 for one focused knowledge use case using agreed materials and users. Answers should point back to sources, and the team needs a correction path. Larger integrations or custom software require separate scope and pricing.
4. Spreadsheets beside the system of record
You pay for a CRM, practice management system, or ERP, but the team keeps another spreadsheet to make the day work. It might track exceptions the main system cannot express, or it might compensate for a field nobody trusts. The sheet is not automatically the problem; it is evidence of a gap worth understanding.
Connecting it before understanding that gap can spread conflicting values faster. A sync does not decide which record is authoritative, who may change it, or what should happen when the two disagree.
What we do first: Trace the workflow with the person who operates it. Follow one item through both systems, record duplicate work and exceptions, and identify the source of truth. Then Choose whether to configure the existing tool, simplify the process, build a connection, or leave the sheet in place with clearer ownership.
The work around the model matters. A clean handoff comes before a new automation.
5. Overlapping subscriptions
The shop has several tools that can draft messages, summarize meetings, or hold customer information. Each arrived for a reasonable purpose. Now the team switches between them, pays for unused seats, and cannot tell which output belongs in the working record.
Adding another assistant can increase that confusion. Consolidation can also fail if it removes a feature someone quietly relies on. The goal is a simpler way to do the job, not a smaller tool count at any cost.
What we do first: list each tool’s actual job, owner, usage, cost, and data handoffs. Apply the money, time, and customer filter. Buy, simplify, and leave alone are all valid outcomes. When the choices are contested or the dependencies unclear, the optional Margin Scan is $1,500 for seven business days. It supports a ranked decision; it is not required when the answer is already clear.
6. Manual reporting
A manager spends Monday collecting exports from the CRM, advertising account, and finance system. The report arrives, then someone asks why two totals disagree. The requested solution is an automatic dashboard, but the real work may include reconciling definitions and exceptions.
Automating the pull without that understanding produces a faster disagreement. A changed field, expired permission, or missing period can quietly turn a plausible chart into the wrong basis for a decision.
What we do first: Trace one reporting cycle and Choose the measure that matters. Define its source, calculation, reporting period, reviewer, and behavior when data is missing. Then scope Implementation for the necessary connectors and checks. Connector work is priced for the agreed system; the $3,500 knowledge-use-case entry price is not a blanket price for reporting integrations.
The named operator should be able to trace a figure to its source and recognize an incomplete run. Leadership keeps judgment over the conclusions, especially when financial or other consequential decisions depend on the report.
7. Tried before; didn’t stick
An automation worked in a demonstration, then broke when a field changed. Or it continued running while the team returned to its old process. The person who built it moved on, and nobody knows whether to repair it or start again.
Rebuilding the same idea without examining the failure can repeat it. The missing piece may be technical, but it may also be an unclear owner, an awkward review step, or a workflow that never fit the operator’s day.
What we do first: examine the existing system and design for Prove. Use representative cases, a baseline, a runbook, and a named operator who can understand, use, verify, override, and recover. A material repair is scoped as implementation; a new subscription does not substitute for fixing the foundation.
When something already works, Ongoing Partnership starts at $1,500 per month for two 45-minute working sessions and one lightweight skill or workflow improvement each month, with a three-month opening term. Larger builds and software costs are separate. Read What an Ongoing AI Partner Actually Does for a Small Business.
8. Want guardrails / human approval
The team wants help preparing work while keeping control over what happens next. A law firm may want a structured intake draft, a dental office a reviewed administrative response, or an estimator a first pass that still requires checking parts and price.
“Someone will check it” is incomplete unless the process names that person, shows them the supporting information, and lets them stop or correct the action. A review step that appears after a message has been sent is too late.
What we do first: define what the system may draft, what it may access, who approves the next action, and how the team recovers from an error. Customer-facing messages remain human-reviewed. We do not recommend a first move that makes autonomous legal, medical, credit, hiring, safety, or financial decisions.
The Sprint can help establish those working habits; implementation must carry them into the system. For an existing workflow, ongoing support can help maintain them as tools change. Human judgment is part of the design in every path.
Choose one useful first move
If three or more of these sound like your shop, a free introductory call can help sort them into a first step. That number is a prompt for conversation, not a qualification threshold. You can also come with one clear problem or an existing solution ready to scope.
Start the intake and tell us what already feels true. See Sprint, Implementation, and Partnership for current scope and pricing. For how the work proceeds, read Trace → Choose → Build → Prove.
Your first engagement should leave you with something useful. The Sprint’s workflows, instructions, and opportunity map stay with your team even if no build follows. A deeper Scan remains optional. We can also recommend a process change, an existing tool, or leaving a task alone.
FAQ
Do we need all eight fixed before we start? No. These are questions to investigate, not prerequisites. Start with a useful task and understand the dependencies that affect it. Some issues may matter now; others can wait.
Sprint or Margin Scan when we picked “all of the above”? Choose the Sprint when the immediate goal is helping a team use AI on real tasks. Consider the optional Scan when an operating area needs deeper analysis before a decision. A free introduction helps establish which question matters first; you do not need both by default.
Is a failed prior automation a reason to wait? It is a reason to inspect the failure. We need to understand the data, ownership, review path, and operator experience before recommending repair, replacement, or retirement. A new build is not the automatic answer.
Will you let AI send customer messages without a human? Customer-facing drafts remain human-reviewed. The workflow names the reviewer and approval point, and consequential decisions remain with qualified people. The controls must work in daily use, not just appear in a policy document.