Summary

For owners deciding whether a business task needs custom work, this guide connects the definition with a practical fit and cost comparison. It explains platform reuse, scoping, controls, maintenance and how to compare complete proposals.

What you'll learn

  • What makes a business workflow custom
  • How to test approval, information and delivery fit
  • What to include in a scoped build and acceptance
  • How to compare full costs over the same period
  • Who maintains the task and handles changed requirements

Say an HR consultancy wants help with client intake. The task needs a particular review sequence, fields from its client records and delivery to its own portal. Before commissioning anything, the team should test whether an existing tool can handle those requirements. If it cannot, the missing work becomes a concrete scope for a custom agent.

What is a custom agent?

A custom agent is an AI system built or adapted around a particular business task: its information, tools, decisions, approval steps and outputs. It can reuse an existing agent platform. “Custom” describes the work done for your business, not a requirement to build every component from scratch.

You usually do not need to train a new AI model to handle a business workflow. The implementation selects suitable existing AI and connects it to the task. Anthropic’s agent-design guidance recommends starting with the simplest solution that works and adding complexity when it improves the outcome.[1]

A standard OpenClaw setup or Hermes setup may be part of a tailored implementation. Both have extension mechanisms: OpenClaw documents plugins.[2] Hermes documents reusable skills and custom workflows.[3] Both can therefore be adapted rather than treated as sealed products.

The useful question is what the delivered system does for your team. Can it prepare the intake record, send the right items for review and stop when information is missing? Who maintains those behaviours as your work changes? The platform name alone does not answer that.

Comparison of configuring an existing tool and adapting or building a custom workflow, covering starting point, fit, operating cost and responsibility.
Compare the actual task and responsibilities. Existing platforms can also support a custom implementation.

What makes an agent custom?

The custom work lies in how the system handles your task. That includes reading the right information, interpreting it, preparing an appropriate output and applying the permissions your business has agreed. Existing and tailored systems can use the same AI, but their surrounding behaviour can differ.

Ask to see your actual task demonstrated, including a difficult case. A bespoke label is less useful than evidence that the system handles your information, review sequence and delivery requirements—and that someone owns its upkeep.

ComponentExisting tool configurationTailored implementation
The AI usedChoose from supported optionsChoose and test suitable supported AI
Tool connectionsUse available connections and permissionsAdapt or build missing connections; verify access
Workflow decisionsConfigure available rulesImplement the decisions your task needs
Approval controlsTest the product’s supported review pathImplement and test additional review requirements
Output formatCheck available fields and formatsBuild the required format and test the receiving system
Information handlingCheck supported structures and accessMap your records and validate missing or inconsistent information

For an intake task, this could mean matching an enquiry to the right client, collecting required documents and preparing a record for a reviewer. The system should handle uncertainty explicitly rather than guess which client was meant. A new connection or field mapping is custom work even if most of the implementation reuses a platform.

Do not assume all products use identical models or expose identical capabilities. Select and test the complete setup against representative work. A polished demonstration with clean examples is a starting point; conflicting records, missing information and a declined approval matter too.

Custom or off-the-shelf: how do you decide?

Choose the option that meets the task’s requirements at an acceptable full cost. A working configuration may be enough. A missing connection, review step or output format may justify adaptation or a custom build. The number of gaps is a review aid, not a price rule.

SignalExisting tool may fitAdaptation or custom work may be needed
Approval flowSupports your exact sequence and rejection behaviourRequired decision or control is missing
Output destinationCan deliver to the actual receiving systemNeeds a new connection or delivery method
Information formatHandles your records and missing valuesNeeds mapping, validation or transformation
Tool stackSupports the actions with suitable permissionsA needed operation or access boundary is missing
WorkaroundsFew, understandable and ownedSeveral dependencies without clear upkeep
Time horizonConfiguration and testing fit your delivery needAdditional development and acceptance work are required
BudgetFull operating proposal fits expected benefitCompare a scoped build with the cost of adaptation
The business-specific work is in the task, connections and controls around the AI.

A small amount of code does not automatically make custom cheaper. It may be an efficient way to adapt an existing tool. Conversely, a product with many familiar connections may still lack the action you need. Compare the delivered behaviour, support and exit route rather than treating “buy” and “build” as mutually exclusive boxes.

Keep a record of each gap: requirement, current limitation, proposed bridge, owner and test. Use this custom-versus-standard comparison alongside your proposal to assess requirements and costs together.

What are the three main fit checks?

Check the approval sequence, information format and output destination against real cases. Each can turn a plausible demonstration into a process your team cannot safely use. A product may handle several review stages or custom fields; test the actual capability rather than assuming a fixed limit.

Approval sequence. Say a client-facing message needs team review followed by partner sign-off. Confirm what happens on approval, rejection, edits and an unanswered request. Make sure the business action waits for the required decision. OpenClaw’s execution approvals concern command execution; they should not be assumed to cover every business message or record change.[4]

Information format. Compare the records the task receives with what the tool expects. Check field names, dates, missing values, duplicate clients and conflicting information. A maintained transformation may solve the gap. Include its upkeep rather than treating it as free glue.

Output destination. Demonstrate delivery to the system the team actually uses. An entry on a connection list does not tell you whether it can perform your required operation with the right access. Check failed delivery, repeated attempts and the receiving system’s acceptance of the output.

A poor fit and a software defect are different problems, and either can occur. Establish which behaviour the product supports, reproduce the failure and agree whether the answer is configuration, a supported fix, an adaptation or a different implementation.

Fit signalMatchedGap to investigate
Approval sequenceActual review, edits, rejection and delivery are supportedA required decision or enforced waiting step is missing
Information formatReal records arrive in a usable, validated shapeMissing mapping, validation or inconsistent-record handling
Output destinationCorrect output reaches the required systemA required operation, format or permission is unavailable

You can count gaps to organise the discussion. Zero gaps still needs testing and cost review; one serious missing approval can matter more than several minor formatting changes. For each gap, compare the proposed remedy and the work needed to maintain it.

Three fit checks: actual approval sequence, information format and output destination; a gap requires a scoped remedy rather than an automatic custom-build verdict.
Severity, supported remedies and operating responsibility matter more than the number of mismatches.

What does a custom agent build involve?

A build connects the task’s information and actions with tested decisions and controls. Agree the expected result and acceptance cases before enabling real changes. The deliverable is a usable business process, together with responsibility for its operation.

Foundation. Select the AI against representative inputs and the quality, speed and cost the task needs. A simpler setup may be enough. Avoid adding several agents when one well-defined preparation step answers the requirement.

Connections. Define what the system may read, change or send. For a recruiting workflow, separate reading candidate information from changing a status or contacting a person. Verify the actual permissions instead of relying on an instruction to behave carefully.

Logic. Decide what starts the task, which information it needs, what it produces and how uncertainty is handled. Include duplicate events and incomplete records. A clear decision path helps the team understand why a case stopped.

Control. Define review, escalation, records of completed actions and a way to stop further work. Test a declined approval and an unavailable receiving system. The custom design still has constraints: connected tools, available permissions, budget and error handling all shape what it can do.

Four parts of a tailored implementation: suitable AI, tool connections, task decisions and tested controls.
Existing components can be reused; each part must fit the agreed task.

How do you scope a custom agent build?

Bring the current process, recent examples and the actions the system may take. These give a supplier a concrete basis for a proposal. Missing requirements can still emerge during testing; a documented process helps manage the work without guaranteeing a fixed deadline or no scope changes.

1

Document the workflow

Describe what arrives, what people do, which decisions they make and where the accepted output goes. Include work before and after the proposed agent task.
2

List required connections and actions

Name the tools, records and fields involved. Separate reading, changing and sending. Ask the supplier to confirm what those tools actually allow.
3

Define review and escalation

Mark the decisions that require approval and the cases that should go directly to a person. Include rejection, changes after review and unanswered requests.
4

Collect difficult cases

Include missing documents, ambiguous clients, repeated enquiries and an existing complaint. Agree which cases must be handled before launch and which remain manual.
5

Agree measurable acceptance

Define accepted outputs, missed cases, corrections and review effort. Set thresholds for this task with the responsible team; do not inherit a universal approval percentage or waiting period.

Ask for assumptions and exclusions in the proposal. A portal connection may be feasible but require access from its owner. A review sequence may require a person to answer promptly. Make those dependencies visible before agreeing a delivery date. What a full agent implementation involves covers the wider handoff.

What does custom agent maintenance involve?

Maintenance keeps the task useful when requirements, connected tools or inputs change. Both an existing platform and a custom implementation need someone responsible for the complete workflow. A vendor may maintain its product while your specific configuration and bridges remain your responsibility.

Instructions and requirements. Revisit the behaviour when service terms, language or output requirements change. Test old and new cases before expanding authority. There is no universal schedule that makes every workflow current.

Connections. Changes in fields, permissions or tool behaviour can affect delivery. Confirm who notices a broken connection, who investigates it and how the team handles pending cases. A platform update does not by itself test your whole process.

New exceptions. Record recurring cases the original design did not handle. Decide whether to extend the task or keep those cases with people. Include review time and repeated corrections in the cost calculation.

Illustrative model-use cost of $24 after 12 months and $48 after 24 months at constant stated volume, plus separate quoted setup, upkeep and other operating costs for either approach.
The model-use subtotal rises with usage. Setup, support and human work must be added; neither approach has guaranteed flat maintenance.

The chart continues the same curve at the stated volume: $72 at 36 months.

How much does a custom agent cost?

Get a scoped proposal for the actual task and compare complete costs over the same period. Include discovery, implementation, tool and AI usage, hosting, testing, review, support and exit arrangements. A tool fee alone is not the cost of an operating workflow.

Cost componentAmount to obtainWhat to include
Scoping and briefQuoted discovery amountRequirements, examples and feasibility; check whether credited
Build: one workflowScoped implementation quoteConnections, decisions, controls and acceptance
Build: several workflowsQuote for shared and separate workShared access plus additional tasks and dependencies
Infrastructure and AI in year oneMeasured usage and provider pricesHosting, model calls, paid tools and repeat attempts
Maintenance from year twoSupport scope and quoteRequirement changes, connection repairs and exceptions
Total year oneSetup plus all first-year running workInclude internal review and implementation effort
Total later yearsRecurring work plus planned changesInclude support, usage, review and any new development

Ask what happens when the task stops, a tool changes or the supplier hands it over. Include ownership of configuration, records and necessary access. The implementation cost breakdown provides a fuller costing method; pricing explains Yardwork’s scoped offers.

How do you compare costs over 24 months?

Compare the same accepted task, volume and responsibilities over 24 months. Use actual proposals for setup and support, current provider rates for usage and measured human work. Count bridges as separate responsibilities, not as automatic evidence that custom is cheaper.

Here is an illustration of the calculation, not a build quote. As of October 2026, Anthropic lists standard first-party Claude Haiku 4.5 input at $1 and output at $5 per million tokens.[5] Assume 10,000 billed input tokens and 2,000 output tokens per task, with 100 tasks monthly. The model subtotal is $0.02 per task, $2 monthly and $48 over 24 months. The volume is hypothetical; the rates are the sourced inputs.

Let O be the actual existing-tool setup quote, B the custom-build quote, T monthly platform charges beyond model usage, R other monthly costs including hosting, support and valued human review, and W the total quoted or measured workaround work over the period. Obtain separate values for each option; avoid counting the same support or subscription twice.

Fit gapsExisting tool: 24-month illustrationCustom: 24-month illustrationDecision
NoneO₀ + $48 + 24 × (T₀ + R₀)B₀ + $48 + 24 × (T꜀ + R꜀)Compare total cost and acceptance evidence
OneO₁ + W₁ + $48 + 24 × (T₁ + R₁)B₁ + $48 + 24 × (T꜀ + R꜀)Assess the particular gap and remedy
TwoO₂ + W₂ + $48 + 24 × (T₂ + R₂)B₂ + $48 + 24 × (T꜀ + R꜀)Compare remedies and ongoing responsibility
ThreeO₃ + W₃ + $48 + 24 × (T₃ + R₃)B₃ + $48 + 24 × (T꜀ + R꜀)Choose a tested, affordable complete option

The same $48 appears only because this illustration assumes equal billed model usage. A real implementation may make more calls or supply longer inputs. At twice the hypothetical volume, that subtotal is $96; at five times, $240. Recalculate from billed usage, including repeated attempts. This illustration excludes caching, batch discounts, paid tool calls and taxes; include applicable charges in your actual comparison.

The subscription is only one line in the cost of a working task.

For workaround work, separate initial development from repairs and ongoing handling. Use quoted development amounts, observed repair effort and recorded recurring hours valued at your actual rate. Do not assume a universal repair frequency or multiply a guessed hourly rate into a supposedly typical price.

Cost comparison separating a sourced $48 model-use illustration over 24 months from actual setup, workaround, support and review amounts for each option.
Both totals need real proposals and measured work. Gap counts do not determine a winner.

What do existing tools do well, and how long does setup take?

An existing tool can reduce development when its supported behaviour fits the task. It still needs configuration, suitable permissions, testing and an operating owner. A tailored implementation adds the work needed for missing behaviour; either route can uncover dependencies that affect the schedule.

OpenClaw’s plugins allow extensions.[2] Hermes’s skills support reusable workflows.[3] Neither feature guarantees fit for every business. Ask for the proposed setup to demonstrate your task. A familiar tool name is not a promise of ready-made support for every operation in Gmail, Slack, Notion or a client portal.

MilestoneExisting setupTailored implementation
Go-liveAfter configuration, access and acceptanceAfter required development, access and acceptance
First output reviewReview representative configured outputsReview representative outputs during development
First complete runTest the whole task before enabling actionsTest the whole task before enabling actions
Gaps become visibleDuring realistic tests and operationDuring realistic tests and operation
Maintenance beginsAgree responsibility before launchAgree responsibility before launch
ReassessmentReview changed requirements and observed frictionReview changed requirements and observed friction

Starting with an existing tool can help gather task evidence when it already meets essential controls. It is not necessary to run a poor fit for a fixed number of months. If it cannot meet a required permission or review boundary, resolve that before allowing the business action. Ask the supplier for a schedule tied to access, development and acceptance work.

What do the choices look like in practice?

Consider three hypothetical businesses. Each should test the proposed task and compare its full cost. The examples illustrate decision paths; they are not client results or measured savings.

A marketing agency. Its proposed weekly account report uses available connections and an existing review step. Test the actual figures, account separation and delivery. If those work and the full operating cost fits, configuration may be enough. Reconsider if a new client requires a different format.

A compliance consultancy. Intake needs team review, partner approval and delivery to a private portal. First check what the selected product supports. A missing step might be handled by a supported adaptation; an unsuitable access boundary might require different development. Compare a complete remedy with a custom proposal rather than declaring a winner from three gaps.

An HR consultancy trying an existing setup. Its pilot reveals repeated manual field mapping. Record the handling time, errors and responsibility. Those observations can inform a bridge or a tailored task. Include migration and handover work in the comparison. Real examples improve the brief without eliminating the possibility of new requirements.

How do you make the final build-versus-buy decision?

Compare acceptance evidence and full operating cost for the options that can meet essential requirements. Do not choose a cheap option that cannot perform the required review or action. Keep the task narrow enough to demonstrate, price and hand over.

Decision sequence: define the task, test essential fit, compare complete proposals and operating cost, then agree acceptance and upkeep.
Resolve essential requirements first; compare the full costs of workable options.

Write down the requirement, what the existing option can do, the proposed remedy and the responsible person. Ask each supplier to demonstrate the difficult cases and identify what remains manual. Compare setup, support, usage, review and exit costs over the same period.

Choose configuration when it provides the needed behaviour and responsibility. Choose adaptation or custom work when it closes a specific valuable gap at an acceptable cost. If neither proposal works, narrow the task or keep it manual. How to build a custom agent gives the next implementation steps.

Frequently asked questions

What is a custom AI agent?

A custom agent is built or adapted around your business task, information, tools, decisions and controls. It can reuse an existing platform; you do not have to build every component or train a new model. OpenClaw and Hermes can support tailored implementations through their extension mechanisms. Custom describes business-specific work, not necessarily software built from scratch. Test the chosen platform’s actual actions and controls.

When should a business choose custom work or off-the-shelf?

Choose adaptation or custom work when it closes a specific valuable gap in permissions, review, information handling or delivery at an acceptable full cost. Demonstrate the task and compare complete proposals; a gap count does not choose the winner. Use an existing configuration when it meets essential task requirements and has acceptable full operating cost and responsibility. Verify the actual review sequence, records and output rather than assuming a universal coverage percentage.

How do custom and off-the-shelf agent costs compare?

Obtain a scoped proposal including setup, AI and tool use, hosting, testing, review, support and exit. Compare the same period and accepted task. The article’s sourced model-use illustration is a subtotal, not an implementation price. It becomes cheaper when its complete quoted and measured cost for the accepted task is lower over the same period. Include development, migration, support, usage and human work on both sides; there is no universal price or mismatch threshold. Compare actual setup and support proposals, current provider rates and measured human work over 24 months. Add workaround development, repairs and recurring handling without counting the same work twice. Use the same accepted task and volume.

How long does it take to build a custom agent?

Ask for a schedule tied to access, required development and acceptance tests. Both routes need realistic testing. A fixed number of days or weeks cannot cover every task, connection and review dependency.

What does custom agent maintenance involve?

Maintain requirements and instructions, tool connections and recurring exceptions. Agree who detects failures, handles pending cases, repairs changes and tests updates. There is no universal maintenance percentage or schedule.

Can a business start with off-the-shelf and switch to a custom agent later?

Yes. Record observed limitations and human effort, then compare a supported adaptation with a tailored implementation. Include migration and handover. You need not wait a fixed number of months or accept missing essential controls.

What are the three signals that predict a poor off-the-shelf fit?

Review the actual approval sequence, information format and output destination. Investigate severity, supported remedies and upkeep. One missing control may matter more than several minor formatting gaps.

Notes

  1. Anthropic, Building effective agents. Accessed October 14, 2026.
  2. OpenClaw, Plugins. Accessed October 14, 2026.
  3. Nous Research/Hermes, Skills System. Accessed October 14, 2026.
  4. OpenClaw, Exec approvals. Accessed October 14, 2026.
  5. Anthropic, Claude pricing. Accessed October 14, 2026.