Skip to main content
AGENCYSCRIPT
CoursesEnterpriseBlog
đź‘‘FoundersSign inJoin Waitlist
AGENCYSCRIPT

Governed Certification Framework

The operating system for AI-enabled agency building. Certify judgment under constraint. Standards over scale. Governance over shortcuts.

Stay informed

Governance updates, certification insights, and industry standards.

Products

  • Platform
  • AI Scripts
  • Certification
  • Launch Program
  • Vault
  • The Book

Certification

  • Foundation (AS-F)
  • Operator (AS-O)
  • Architect (AS-A)
  • Principal (AS-P)

Resources

  • Blog
  • Agency Archetype Quiz
  • Free Live Training
  • Build AI Agents Masterclass
  • Build with AI Challenge
  • OS Plugin Install
  • Verify Credential
  • Enterprise
  • Partners
  • Pricing

Company

  • About
  • Contact
  • Careers
  • Press
© 2026 Agency Script, Inc.·
Privacy PolicyTerms of ServiceCertification AgreementSecurityCookies

Standards over scale. Judgment over volume. Governance over shortcuts.

On This Page

Scoping ChecksThe Job Is One SentenceSuccess Is Defined Before BuildingPlatform ChecksCapability and CeilingPortabilityGrounding ChecksKnowledge Is Connected and VerifiedHonest Fallbacks ExistShaping ChecksRole, Tone, and BoundariesThe Failure Path Is DesignedTesting ChecksReal and Hostile InputsLaunch ChecksStaged Rollout and LoggingA Path for Bad AnswersOperating ChecksOngoing Review and RefreshIntegration and Action ChecksActions Work and Fail SafelyConnections Are Verified End to EndSecurity and Privacy ChecksData and AccessFrequently Asked QuestionsHow should I actually use this checklist?Which checklist items matter most?Why is verifying retrieval a separate checklist item?Should I run the whole checklist at once?What is the most overlooked checklist item?Key Takeaways
Home/Blog/What Belongs on a 2026 Conversational Build Review List
General

What Belongs on a 2026 Conversational Build Review List

A

Agency Script Editorial

Editorial Team

·April 10, 2016·8 min read
ai chatbot development platformsai chatbot development platforms checklistai chatbot development platforms guideai tools

A checklist is only useful if you can actually run it against a project and get a verdict. This one is built to be used: each item is concrete enough to mark pass or fail, and each carries a one-line justification so you know why it matters and can skip it deliberately when it does not apply. Run it before launch and again on a cadence afterward, because chatbot quality drifts.

The list is organized by phase, scoping, building, grounding, testing, launching, and operating, because the right questions change as a project moves. An item that is critical at grounding is irrelevant at scoping, and trying to check everything at once produces a list nobody runs. Work it in order, and treat any unchecked item as a known risk you have chosen to accept rather than one you forgot.

Scoping Checks

Get these wrong and nothing downstream can save the project.

The Job Is One Sentence

  • The bot has a single, written, one-sentence job. A bot that needs a paragraph to describe its job is too broad to be good at any part of it.
  • It is explicit about what it will refuse or escalate. Boundaries are a feature; an undefined boundary becomes a confident wrong answer.
  • Your error tolerance is stated. Knowing whether occasional wrong answers are acceptable tells you how much to invest in grounding and testing.

Success Is Defined Before Building

  • You have a measurable definition of a successful conversation. Without it, you cannot tell whether the bot works or just runs. The scoping logic is detailed in Inside One Support Team's Move to a Conversational Assistant.

Platform Checks

The tool should fit the job and not trap you.

Capability and Ceiling

  • The platform can do the job at the simplest tier you need. Paying for power you will not use is waste; lacking power you will need is worse.
  • You know the platform's ceiling. Knowing where it stops prevents the silent outgrowing that strands successful bots.

Portability

  • Your instructions, knowledge, and conversation data can be exported. Portability is your only clean exit if the tool becomes a cage, a trade-off weighed in Hard-Won Defaults for Teams Building Conversational Software.

Grounding Checks

This is where accuracy is won or lost.

Knowledge Is Connected and Verified

  • The bot is grounded in your real, current documents. Fluency without grounding is confident fiction.
  • You have verified retrieval returns the right document for known questions. Grounding that is configured but unverified is grounding you cannot trust.
  • Documents are structured for retrieval, not just human reading. Poorly structured docs cause retrieval to miss answers that technically exist.

Honest Fallbacks Exist

  • The bot says it does not know rather than guessing. An admitted gap is more useful and less damaging than a confident error.
  • Knowledge has a clear owner and a defined refresh trigger. Grounding decays the moment your business changes, so someone must own keeping it current.

Shaping Checks

The instructions are where behavior is born.

Role, Tone, and Boundaries

  • The bot's role and tone are written down and consistent. An undefined personality drifts and confuses users across conversations.
  • Boundaries and refusals are encoded in the instructions, not just hoped for. A boundary that lives only in your intentions is a boundary the bot will cross.
  • The instructions are readable and version-controlled. Prompts drift over months of edits, and an unversioned prompt becomes something nobody dares change.

The Failure Path Is Designed

  • The bot has a defined way to signal uncertainty and escalate. Failure handling is part of quality, since every bot is sometimes wrong.

Testing Checks

Friendly testing is not testing.

Real and Hostile Inputs

  • You have an evaluation set of real questions with defined good answers. A standing set is the only way to catch silent regressions.
  • You have probed adversarially: off-topic, rude, instruction-leak, and false-promise attempts. Hostile inputs are guaranteed in production, so rehearse them. The patterns are in Walking Through Bots That Shipped and the Ones That Stalled.
  • Escalation rules are tested by triggering them several ways. Untested escalation is the classic source of embarrassing failures.

Launch Checks

Going live should be staged and observable.

Staged Rollout and Logging

  • The bot launches to a small audience first. Real users ask what your test set never covered; learn it cheaply.
  • Conversation logging is on from the first message. You cannot improve what you cannot see, and the test-to-real gap appears immediately.

A Path for Bad Answers

  • Users can flag a wrong answer easily. A flagging path turns failures into improvement signal instead of silent attrition.

Operating Checks

The list you run forever, not once.

Ongoing Review and Refresh

  • Logs are reviewed on a set cadence. Unreviewed logs mean drift goes unnoticed until users complain.
  • Knowledge is refreshed when the business changes. Stale knowledge turns a good bot confidently wrong, as warned in Quiet Ways a Conversational Bot Project Goes Wrong.
  • Cost and latency are tracked as volume grows. Both stay invisible until success makes them painful.

Integration and Action Checks

A bot that cannot do anything is often a bot not worth building.

Actions Work and Fail Safely

  • The bot can take the actions its job requires, and you have tested each one. A bot that only talks frequently fails to beat a well-organized help page.
  • Actions fail safely when something goes wrong. A booking that half-completes or a lookup that returns stale data should degrade gracefully, not silently mislead.
  • The handoff to a human carries conversation context. A handoff that drops the thread punishes the user for trusting the bot.

Connections Are Verified End to End

  • Every system the bot touches, CRM, calendar, database, has been tested with real, not mock, data. Integrations that work in a sandbox often break against production quirks, the kind of gap that stalls otherwise solid projects.

Security and Privacy Checks

The moment real data is involved, these stop being optional.

Data and Access

  • You know where user data goes, how long it is retained, and who can see it. For regulated contexts this is a gating requirement, not a footnote.
  • The bot cannot retrieve or reveal data outside its scope. Access boundaries are part of the build, and the consequences of skipping them appear in Quiet Ways a Conversational Bot Project Goes Wrong.
  • Instruction-leak and manipulation attempts have been tested and blocked. A bot coaxed into revealing its rules or exceeding its access is a liability waiting to surface.

Frequently Asked Questions

How should I actually use this checklist?

Run it against your project before launch and then on a cadence afterward, marking each item pass or fail. Work it in phase order, since the right questions change as the project moves. Treat any unchecked item as a risk you are knowingly accepting, not one you forgot, which keeps the list honest.

Which checklist items matter most?

The scoping items, because everything downstream depends on them. A bot with one clear job and a defined notion of success can pass the rest; a bot without those cannot be grounded, tested, or measured meaningfully no matter how diligent the later phases are. Get scoping right first.

Why is verifying retrieval a separate checklist item?

Because grounding that is configured but unverified is grounding you cannot trust. Teams routinely connect documents and assume it works, then ship a bot that cannot find its own knowledge. Explicitly confirming the right document is retrieved for known questions catches a large share of accuracy problems before launch.

Should I run the whole checklist at once?

No. It is organized by phase on purpose, because an item critical at grounding is irrelevant at scoping. Trying to check everything simultaneously produces a list nobody runs. Work it in order as the project progresses, and revisit the operating section repeatedly after launch.

What is the most overlooked checklist item?

The operating checks, especially knowledge refresh and log review. Teams treat launch as completion, so the items that keep a bot good over time get skipped. A great bot that is never operated decays into a wrong one, which is why these items belong on a list you run forever.

Key Takeaways

  • A checklist is only useful if each item is concrete enough to mark pass or fail; this one is built that way.
  • Work the list in phase order, scoping, platform, grounding, testing, launch, operating, since the right questions change.
  • Scoping items matter most; a clear one-sentence job and a defined success metric enable everything else.
  • Verify retrieval explicitly, because configured-but-unverified grounding is grounding you cannot trust.
  • Probe adversarially and test escalation rules; hostile inputs and untested escalations cause real failures.
  • The operating checks run forever, not once; skipping them is how good bots quietly decay.

Search Articles

Categories

OperationsSalesDeliveryGovernance

Popular Tags

prompt engineeringai fundamentalsai toolsthe difference between AIMLagency operationsagency growthenterprise sales

Share Article

A

Agency Script Editorial

Editorial Team

The Agency Script editorial team delivers operational insights on AI delivery, certification, and governance for modern agency operators.

Related Articles

General

Rolling Out AI Hallucinations Across a Team

Most teams discover AI hallucinations the hard way — a confident-sounding wrong answer makes it into a client deliverable, a legal brief, or a published report. The damage isn't just to the output; it

A
Agency Script Editorial
June 1, 2026·11 min read
General

A Model Behind an API Is Only Potential

Large language models don't do much on their own. A model sitting behind an API is potential, not capability. What converts that potential into something useful—something that drafts, classifies, summ

A
Agency Script Editorial
June 1, 2026·11 min read
General

Case Study: Large Language Models in Practice

Most teams that fail with large language models don't fail because the technology doesn't work. They fail because they treat deployment as a one-time event rather than a discipline — pick a model, wri

A
Agency Script Editorial
June 1, 2026·11 min read

Ready to certify your AI capability?

Join the professionals building governed, repeatable AI delivery systems.

Explore Certification