Iron Goo
Two travel agency staff study a colored responsibility grid on a glass wall in a warmly lit office.

How to Structure a Small Team When AI Does the Work

Atamyrat Hangeldiyev
Atamyrat Hangeldiyev
Systems Architect
March 23, 2026
On this page
Business

At the hypothetical 15-person Waypoint Group Travel, three trip designers have used AI chat assistants daily for four months. Two use personal accounts, one uses a company card, and a web agency runs the site chatbot. A client spots a museum on an itinerary for a day it is closed. The owner asks who checked the draft and finds that nobody owns the tools, the prompts or the review.

The answer to how to structure a small team is to write one responsibility map for each core workflow. Give every row one named Accountable person, and give AI tools and human review points their own rows.

Waypoint designs and books group tours. Its core workflow runs from the first inquiry to a confirmed booking. The chart uses roles as column labels; the working copy would also put a person's name beside each role.

TaskOwnerOps leadSales leadTrip designerSenior designerSales coordBookingsFinanceWeb agency
1 First reply to web inquiry (chatbot)IACR
2 Qualify inquiryACR
3 AI-assisted itinerary draftAIRCC
4 Itinerary review before client sees it (review point)AR
5 Price and margin sign-offACCR
6 Send quoteARI
7 Supplier bookings and depositsAIRC
8 AI tools: seats, cost, data rules, prompt libraryARCCCCI

A RACI chart is one type of responsibility assignment matrix. As PMBOK Guide based exam material puts it, the letters mean:

  • Responsible, R: “Those who do the work to achieve the task.”
  • Accountable, A: “the one ultimately accountable for the correct and thorough completion of the deliverable or task”.
  • Consulted, C: “Those whose opinions are sought; and with whom there is two-way communication.”
  • Informed, I: “Those who are kept up-to-date on progress, often only on completion of the task or deliverable”.

The difference between Responsible and Accountable is the difference between doing the work and owning its completion. A task can have more than one R, but the chart assigns exactly one A to each task or deliverable. At Waypoint, the trip designer drafts the itinerary in row 3, while the operations lead is accountable for that draft.

Responsibility map highlighting one accountable owner per row and separate AI tool and human review rows

Each workflow row has one accountable owner. The AI tool and human review rows make two easily missed responsibilities visible.

The owner's column: rows 5 and 8

Waypoint's owner is A on two rows: price and margin sign-off, and AI tool governance. Those are decisions the hypothetical company has chosen to keep with its owner. In row 1 the owner is only informed about the chatbot's first reply, so routine inquiries do not wait for an owner decision.

Every owner A row costs decision time. The owner needs enough context to judge the work, an opportunity to ask for changes, and a way to resolve exceptions. If ten rows require that attention, the team has ten places where work can wait for the same person.

Count the owner's A cells and ask what decision each one reserves. For a row where a lead can make the decision using an agreed policy, give that lead the A and state when an exception must go to the owner. Keep the owner informed where the information is useful; an I does not require the owner to approve the task.

Row 5 is a useful boundary. Finance prepares the price and margin sign-off, while the owner owns the final decision. Waypoint could later set a price and margin policy that lets another person take the A, but the chart should show the authority the company has actually given them.

Workflow owners: one A on every Waypoint row

Can a RACI have two accountable people? The single-accountable rule assigns one A per task or deliverable. Two A cells can create circular approvals when each person expects the other to make the call.

A row with no A leaves nobody explicitly accountable for completion. That does not prove nobody will do the work; it means the chart has not settled who owns the result when the work is late, incomplete or disputed. For row 2, Waypoint gives the sales lead that responsibility even though the sales coordinator qualifies the inquiry.

Our rule of thumb is to split a row if you cannot name one A. For example, if one person owns preparing a price and another owns approving a margin, make those separate tasks when their decisions need separate owners. Write the boundary plainly enough that a new employee can tell when one row ends and the next begins.

The same principle applies to measurement. Put one person's name beside each KPI, as you would when setting KPIs and targets that drive action. Other people can contribute data or do the work, while the named owner is responsible for deciding what happens when the measure moves.

AI tool ownership: row 8 and the tool register

Row 8 assigns the operations lead to maintain the tools and the owner to be accountable for them. The AI assistant itself never gets an A: it cannot approve a data rule, pay a bill, review its own output on the company's behalf or decide to stop its use.

Start the register with tools people already use. In the Microsoft and LinkedIn Work Trend Index, a vendor survey of 31,000 knowledge workers across 31 markets, 78% of AI users reported bringing their own tools to work. The reported figure was 80% at small and medium-sized companies. Waypoint's two personal accounts would otherwise be easy to miss if the register began with company card charges alone.

Use one register row per tool. The following fields turn row 8 into work someone can actually perform.

Register fieldWhat Waypoint records
ToolProduct and account or workspace used
Business ownerNamed person accountable for its use
AdminNamed person who manages access and settings
Payment and monthly costWho pays, seat count, billing period and monthly cost
Data allowed and bannedPermitted inputs and explicit exclusions, such as passport numbers and minors' details
Prompt and example libraryLocation of the maintained prompts, examples and instructions
Review point it feedsThe chart row where a person checks its output
Stop switchHow to suspend the tool or workflow and who can do it
Renewal dateDate the owner must review the subscription

A cost line should show the arithmetic. As a receipt read on 30 October 2026, Claude Team pricing lists a standard seat at $25 per month billed monthly, or $20 per seat per month billed annually. Five seats therefore cost $125 a month on monthly billing, or $100 a month billed annually. The register should still name who pays and who reviews renewal.

The NIST AI Risk Management Framework is voluntary guidance. Its GOVERN 1.6 says, “Mechanisms are in place to inventory AI systems and are resourced according to organizational risk priorities.” GOVERN 2.1 says, “Roles and responsibilities and lines of communication related to mapping, measuring, and managing AI risks are documented and are clear to individuals and teams throughout the organization.” A small register and a named owner make those ideas usable without pretending the spreadsheet itself manages the risk.

Human review: row 4 before the itinerary leaves

Waypoint gives review its own row. The trip designer is R for the draft in row 3; the senior designer is R for review in row 4. The operations lead remains A for both, so the reviewer has a person to escalate to when an itinerary cannot be cleared.

The review must include observable checks. For the closed museum, the senior designer checks the proposed visit date against the venue's current opening information before the itinerary goes to the client. The reviewer also records their name on the approval, so a disputed itinerary can be traced to an actual review.

A human signature alone does not prove that errors are being caught. In a study by Dratsch and colleagues involving 27 radiologists, very experienced readers rated cases correctly 82.3% of the time when a purported AI suggestion was correct, versus 45.5% when it was wrong. That is evidence of automation bias in a medical task, not a measurement of Waypoint's travel reviews.

Our rule of thumb is to seed known-wrong items into a review test and measure the catch rate: known errors caught divided by known errors presented. Tell reviewers the test process exists, record the result, and use missed errors to improve the check. The human oversight and guardrails guide sets out seeded checks, approval queues and stop controls.

Name who can pull the stop switch when the output is unsafe or the review queue cannot be handled. At Waypoint, that could mean the operations lead pauses use of the drafting assistant until the issue is resolved. The register records the mechanism; the review row records where the issue should be found.

Generalists and specialists: the people behind Waypoint's columns

In a company of 5 to 25 people, one person may hold several roles. The chart's columns describe roles, so a named person can appear in more than one column. A generalist can handle several kinds of work while the chart still makes each responsibility explicit.

That flexibility has a limit: the row 4 reviewer must be a different person from the row 3 drafter. If the only available senior designer wrote a particular draft, Waypoint needs another qualified reviewer for that draft. Putting the same name in both columns would make the chart look complete while removing the independent check.

Our rule of thumb for a single point of failure is to find any column whose R cells nobody else can cover. For each such column, name a backup and check that the backup has the access, instructions and time to do the work. If the answer is no, the chart has identified a staffing or training decision.

A specialist is useful when a task needs judgment that a generalist cannot reliably supply. A generalist is useful when several smaller duties need an owner and the work fits one person's capacity. The chart does not choose either model automatically; it shows which tasks and review points the people in those roles must cover.

Span of control: Waypoint's owner and operations lead

How many direct reports should a manager have? There is no single number that fits every team. Gallup's January 2026 analysis of US managers puts the mean at 10.9 direct reports in 2024 and 12.1 in 2025, while the median was about five to six.

The distribution matters. Gallup reported that 37% of managers led fewer than five people and 13% led 25 or more. Its assessment was that manager talent trumps span, with four factors to examine: team engagement, the manager's individual contributor workload, manager talent and feedback. Gallup singles out meaningful feedback to each employee at least once a week.

V. A. Graicunas offered a different lens in “Relationship in Organization,” published in March 1933. His formula counts possible direct single, cross and direct group relationships for a manager with n direct reports. It is a maximum case of possible relationships, not a measurement of actual workload.

Possible relationships for n direct reports

n * (2^(n-1) + n - 1)

The count is 44 for four direct reports, 100 for five, 222 for six and 5,210 for ten. If Waypoint's owner directly managed the other 14 people, the formula gives 114,870 possible relationships. If the owner had three direct reports, the count would be 18; an operations lead with seven direct reports would have 490.

Relationship webs show possible working relationships growing as a manager adds direct reports

The web becomes denser as direct reports are added. Graicunas counts possible relationships in a maximum case, not the interactions a manager actually handles.

Graicunas suggested a maximum of five subordinates, probably four in most cases. Fred Nickols calls the choice a judgment call. Those historical suggestions are useful prompts for inspection, not a mandate to reorganize every team at five people.

Our rule of thumb for considering a manager layer is to inspect the work before changing the org chart. Check whether the owner can give each direct report meaningful weekly feedback, whether decisions wait for the owner, and whether a proposed lead has enough time and authority to manage. A new layer only helps if its lead can make the decisions the chart assigns.

Contractors and agencies: the web agency in row 1

A contractor or agency can be R in a RACI chart. Waypoint's web agency runs the chatbot in row 1, while the sales lead is A for the first reply to a web inquiry. That arrangement gives the outside party a clear job and gives the company a named person to address an incorrect answer or a misleading price.

Our rule of thumb is to keep A inside the company for customer-facing or money-bearing rows. Define what the agency may change, what it must send for approval, and how quickly the company must be told about an error. Put the agency's contact name in the working chart so the R is a person who can receive and act on the assignment.

NIST's voluntary AI RMF GOVERN 6.1 says, "Policies and procedures are in place that address AI risks associated with third-party entities." For each outside party, name the person who grants access, the person who revokes it and the offboarding date. Check that the agency's tool and data access are listed in the register.

Iron Goo's operations service follows an audit, build and operate model for chatbots that handle repeat inbound questions and an agentic operations layer for internal operations work, with scoped credentials and audit logging. A company using an outside operator still needs to name its own accountable person on the workflow row.

Changed roles: new work around Waypoint's AI drafts

How do roles change when AI takes over part of a task? The remaining work can shift toward review, exceptions and maintaining the material the tool uses. The time saved and its destination need to be measured in the actual role.

In the March 2026 revision of Humlum and Vestergaard's study, covering 25,000 workers across 7,000 Danish workplaces and 11 occupations, adopters reported saving about 3% of work hours and 85% reallocated saved time. The researchers ruled out earnings effects above 2% two years after ChatGPT. Those findings do not predict the result for an individual travel agency.

An unreviewed Census working paper reported that 2.0% of AI-using firms reported AI-related employment decreases. It also reported no task effect at 52% of AI-using firms; among firms reporting a task effect, about 66% reported augmentation only. These are working-paper findings, not official statistics.

Manager support is associated with use, although association does not establish cause. Gallup reported in February 2026 that, within organizations making AI available, frequent use was 78% among employees who strongly agreed their manager actively supported AI use, compared with 44% among those who did not strongly agree.

Before Waypoint launches a changed drafting process, our rule of thumb is to add or clarify these assignments:

  • Review and release: Name who checks each draft and who can release it to a client. Row 4 covers the review; the release authority must also be clear.
  • Examples and exceptions: Name who maintains approved prompts and examples, records known failure cases and decides when they need revision.
  • Exception queue: Name who handles drafts the reviewer cannot clear, including who contacts a supplier or asks for a new itinerary.
  • Allocated time: Put review, library maintenance and exception work into the relevant roles' schedules. Check the queue against that capacity after launch.

Use the AI change management guide to roll the changed responsibilities out to staff. When the chart exposes a skill or capacity gap, the hiring and skills guide helps turn the revised role into a hiring or training decision.

The one-page version: checks against all eight rows

Copy the workflow rows and role columns onto one page. Put names beside the roles, keep the tool register alongside it, and check the chart against actual practice with the people doing the work.

  • Every task row has exactly one A.
  • Every task row has at least one R.
  • No AI tool is listed as an A.
  • The R for a review is a different person from the R who drafted that item.
  • The owner's A cells are counted, and each represents a decision the owner has chosen to keep.
  • Every AI tool used in the workflow appears in the register with a named business owner and a renewal date.
  • Each contractor has a named company contact for access, a person who can revoke it and an offboarding date.
  • Each role with R cells has a named backup who can perform the work or a recorded plan to create that cover.
Free ebook

Put your next steps in order

Build on what you have just read with a free ebook on semantic SEO and AI visibility. Get worked examples, decision rules, checklists and a do-this-week action in every chapter, plus a six-week plan for your site.

By signing up and confirming your address, you consent to receive the free ebook plus one email every two weeks with 3 or 4 new Iron Goo videos, guides and articles, and nothing else; unsubscribe any time with one click.

Related in Business

Ready to move?

Tell us where your business is today. A written assessment lands in your inbox within two business days.

Talk to us