Handovers That Do Not Lose the Person in the Gap

The operational problem

The most expensive mistake in care is often not malice. It is a gap between two good people.

This brief is for shift leads, key workers and clinical coordinators. It applies the operating approach in A Practical AI Roadmap for NGO Leadership Teams. It does not repeat the flagship’s full workforce context or implementation phases. Its narrower question is whether the next person can understand and act on what matters without treating a fluent summary as proof that information was transferred.

Scope: This is an operational technology briefing on information transfer between people, shifts, and services. It is not clinical guidance. Human verification of critical handover content is mandatory.

In one line: Details die between two good people: the strategy that worked yesterday, the parent preference, the school contact who answers, the risk that is easy to understate when tired.

What a real week looks like

At 16:45, an outgoing worker tries to explain a difficult afternoon while answering a message and finishing notes. The incoming colleague receives a mixture of remembered detail, shorthand and an old case note. The useful calming approach is mentioned late, the parent call is not, and the action that must happen tomorrow sits in an inbox. No one intended to lose the person. The handover had no protected structure or confirmation.

This is an operational technology brief, not clinical guidance. It concerns the transfer of information between people, shifts and services. It must never replace local clinical, medication, safeguarding or professional procedures. Human verification of critical content is mandatory.

What the evidence already says

WHO patient-safety guidance identifies communication during handover as a safety issue, while NHS SBAR material provides a compact structure for consistently communicating a situation, relevant background, assessment and recommendation. These sources are health-sector guidance, not a licence to transplant a clinical tool unchanged into every NGO. Their useful lesson is narrower: a stable skeleton and a read-back make omissions easier to notice.

Eurofound’s social-services research records the background pressure: staff shortages, strained conditions and uneven digitalisation. The Charity Digital Skills Report 2026 adds a relevant caution: 79% of its 807 charity respondents report AI use as of 2026-07, but only 28% have a digital strategy in place and 35% do not trust AI tools. A tool that produces a fluent handover in seconds is not necessarily safer if its speed encourages staff to skip the outgoing person’s confirmation. Adoption does not establish an accountable handover process.

Where technology helped - and where it stalled

Searchable notes, shift checklists and approved templates can reduce the time spent reconstructing recent events. A controlled system may pre-fill dates, known contacts and open tasks from approved records, leaving the outgoing worker to confirm what matters now. This can make a handover more consistent when staff rotate.

It stalls when generated summaries silently omit a risk, combine two people with similar names, or present stale information as current. NIST identifies confabulation and human-AI configuration risks; in handovers, a polished omission is more dangerous than a visibly incomplete template. Technology does not remove the need for verbal confirmation where local policy requires it.

Structure fixes more of this than software does

Most failures in the handover week described here are not AI-shaped. An outgoing worker finishing three tasks while explaining a difficult afternoon is a staffing and scheduling problem; no drafting tool creates the ten protected minutes nobody allocated. A parent call left out of the account is a habit gap in what gets said aloud, not a retrieval gap. Even the SBAR skeleton works because it is a checklist a person follows under pressure, not because it reasons. Gartner’s 2026 review of generative AI project failures found at least half of such projects abandoned after proof of concept, often for poor underlying data quality, a real risk here: a tool built on inconsistent notes will produce confident nonsense faster than a rushed person would.

The pattern worth an agentic layer sits above any single shift. A month of rota data shows which service or time of day produces the most incomplete handovers, a pattern no outgoing worker can see from inside their own Tuesday. A system that reads recurring omissions across many shifts and flags a service to the supervisor before a near-miss becomes a real one is reasoning about where risk concentrates, then prompting a named person, rather than waiting to be asked. McKinsey’s 2025 State of AI survey found most organisations layer AI onto an unchanged process with no material effect, while the minority who redesign around it capture most of the gain. A vendor selling a generic handover assistant without learning that this team’s skeleton and read-back habit already work is offering transcription dressed as a fix.

A safer AI-assisted path

1. Agree one operational skeleton. This step has no AI role: local mandatory fields and headings such as what matters now, what changed, open actions, who to contact and what must be escalated must be agreed by the team, not inferred from old notes. The skeleton is the checklist people follow under pressure; do not assume a generic form or generated summary covers clinical or safeguarding requirements.

2. Protect a short handover window. This step has no AI role: scheduling and staffing must make the incoming and outgoing person available for the critical transfer. A form cannot compensate for an outgoing worker who is simultaneously completing three other tasks.

3. Use assisted pre-fill only from approved records. The system may collect factual fields, but it must show its source and timestamp. No consumer AI should receive identifiable case data.

4. Require spoken confirmation and read-back. This step has no AI role: outgoing and incoming staff confirm critical items according to local policy, and a generated draft is preparation, not proof that transfer occurred. No tool replaces verbal confirmation where local policy requires it.

5. Sample omissions and near-misses. Review a small number of handovers every month, including apparently routine ones. Where an agentic layer has flagged a service or shift with recurring omissions, the supervisor uses that signal to decide whether the skeleton is too long, too vague or ignored under pressure; learn from sampled transfers, not from draft volume alone.

Where humans must intervene

Outgoing and incoming staff both own the transfer. Critical clinical, safeguarding, and medication-related items are verbally confirmed, not only read from a draft. AI never silently drops a risk flag.

The supervisor’s role is also operational: decide when a handover is too complex for the normal path, ensure the right specialist is involved, and make time for escalation. A tool cannot judge whether a reluctance to eat, a change in communication, or a missed contact is clinically meaningful. It can only make the relevant record easier to locate.

The person receiving support should not be lost behind the template. Where appropriate and lawful, their preferences about communication, routine and involvement should be captured in a way the next worker can understand. That is different from making sensitive detail widely available. Information sharing remains bounded by role, necessity and local policy.

Risks and failure conditions

Fluent incomplete handovers, skipped read-back, consumer tools holding identifiable detail and drafts treated as the record of truth are predictable failures. Another is false standardisation: a template becomes so long that staff click through it, while the actual change in the person’s needs is buried in free text.

A protected ten-minute handover window with an agreed skeleton of headings costs nothing to trial and will show within a week whether the gap is time, structure or both, before any drafting tool enters the conversation.

A ninety-day pilot boundary

For the first ninety days, use one shift type or one clearly bounded service. Keep the existing governance route and introduce only the agreed skeleton, a protected transfer window and a small audit sample. Do not change medication, clinical decision-making, safeguarding escalation or statutory documentation through this pilot.

At day 30, inspect whether staff use the headings and whether incoming staff need fewer follow-up messages. At day 60, review omissions and near-misses with the people doing the work. At day 90, decide whether the process is clearer and safer enough to retain. The decision should be based on sampled transfers, not on the number of drafts generated.

Measures that matter

• Use of the agreed skeleton. Sample the completed handover and the live transfer separately. A populated form does not establish that the receiving worker understood the priority.

• Critical-item confirmation rate. Check locally defined critical fields and read-back requirements, not a generic completeness score. Investigate missing confirmation without treating the sample as a disciplinary trap.

• Information-gap near-misses. Record whether an omission, ambiguity or stale note contributed to a near-miss. The aim is to improve transfer, not to use AI as a convenient culprit.

• Time and completeness together. A shorter handover counts only if required items remain complete and receiving staff report that they can act without chasing the previous shift.

Decision questions

1. Which locally defined items require spoken confirmation and read-back, rather than a completed field alone?
2. Where is the authoritative current record, and how will pre-fill show its source and timestamp?
3. Who decides when a transfer is too complex for the normal handover route?
4. What identifiable information is prohibited from any consumer drafting tool?
5. At day 90, will sampled handovers show fewer information gaps without shorter, weaker conversations?

FAQ

Who is this brief for?

shift leads, key workers, and clinical coordinators, and colleagues who share the same week. For strategy across workflows, use the NGO AI roadmap.

Is this clinical or safeguarding guidance?

No. See the scope note above. Follow your professional standards and local policy.

Can AI summarise a handover?

It may prepare an approved draft from approved records, subject to local governance. It must not replace outgoing-person confirmation, incoming-person read-back or specialist judgement.

What is the first step with no budget for a platform?

Agree the operational skeleton from step one on paper and protect a fixed transfer window for one shift type, then sample a fortnight of handovers against it. If omissions still slip through a properly protected, structured handover, the gap is process discipline, not a missing tool, and that has to be fixed first regardless of what gets bought afterwards.

References

1. David Saliba, A Practical AI Roadmap for NGO Leadership Teams, AIMonger (2026). https://aimonger.com/whitepapers/ngo-practical-ai-roadmap-leadership/
2. Charity Digital Skills Report 2026 (807 respondents; launched 9 July 2026). https://charitydigitalskills.co.uk/report/
3. NIST, AI Risk Management Framework. https://www.nist.gov/itl/ai-risk-management-framework
4. Regulation (EU) 2016/679 (GDPR). https://eur-lex.europa.eu/eli/reg/2016/679/oj
5. Regulation (EU) 2024/1689 (EU AI Act). https://eur-lex.europa.eu/eli/reg/2024/1689/oj
6. World Health Organization, Patient Safety Solutions / communication during patient handovers (quality of care guidance hub). https://www.who.int/teams/integrated-health-services/quality-of-care-and-patient-safety/patient-safety-guidance-and-tools/patient-safety-solutions
7. NHS England, SBAR communication tool (situation, background, assessment, recommendation). https://www.england.nhs.uk/improvement-hub/wp-content/uploads/sites/44/2017/11/SBAR-Implementation-and-Training-Guide.pdf
8. OECD, Artificial Intelligence and the health workforce (2024). https://www.oecd.org/content/dam/oecd/en/publications/reports/2024/11/artificial-intelligence-and-the-health-workforce_c8e4433d/9a31d8af-en.pdf
9. HCPC, Our expectations for your record keeping. https://www.hcpc-uk.org/standards/meeting-our-standards/record-keeping/our-expectations-for-your-record-keeping/
10. Eurofound, Social services in Europe: Adapting to a new reality (2023). https://www.eurofound.europa.eu/en/publications/all/social-services-europe-adapting-new-reality
11. Gartner, “Why Half of GenAI Projects Fail: Avoid These 5 Common Mistakes” (2026). https://www.gartner.com/en/articles/genai-project-failure
12. McKinsey & Company, “The State of AI: Global Survey” (2025 edition). https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai