How a small business actually uses agentic AI to save time, and what still needs human hands
When a job that once took a day drops to 15 minutes, the savings are obvious, and so are the trade-offs. David Gewirtz, reporting for ZDNET, walked five everyday small-business tasks through an agentic workflow (ChatGPT Work, with occasional help from Anthropic’s Claude and OpenAI’s Codex) and published practical outcomes: meaningful time savings, useful, but imperfect outputs, and a consistent requirement for human review.
“ChatGPT Work can tackle real small-business chores.”
“The AI saved me time across five very different tasks.”
“The results can be useful, but still need human review.”
Executive summary
- Where agents help most: data aggregation, email triage, initial product discovery, hardware shortlists, and first-pass financial/contract checks.
- Primary risks: hallucinations, misattributed names/roles, visual polish errors, and connector-permission exposure.
- Three practical first steps: run one narrow, read-only workflow; enforce immediate revocation of connector access; add a short human verification checklist before actioning outputs.
What “agentic” means here
“Agentic” tools can chain actions and interact with external services (email, files, web) to complete multi-step workflows instead of returning a single reply. Example: fetch emails, extract invoices, aggregate amounts, and export a spreadsheet, all in one run.
Five real workflows and what they taught
1) Email scanning and summarization
What he did: Enabled a Gmail connector for a focused task, asked the agent to summarize the last few messages in a thread, then immediately disabled the connector.
“Summarize the last 5 messages from OpenAI’s PR people and capture decisions, open questions, and what I should follow up on next.”
What the agent produced: Concise decision and action lists and follow-ups, useful for triage.
What went wrong and how to fix it: The first prompt missed communications routed through an external PR agency. Takeaway: be explicit in prompts, specify addresses, aliases, or domains, and validate by spot-checking threads. Use narrow-scope OAuth tokens or temporary accounts rather than broad admin credentials.
2) Dataset → slide (aggregation, charting, export)
What he did: Fed a San Francisco public “Fire Incidents” dataset (≈750, 000 rows) to an agent and asked for a PowerPoint slide showing the top 10 battalions by incident count.
What the agent produced: Aggregation and chart creation completed in roughly 13 minutes, and a presentable slide was exported automatically.
What went wrong and how to fix it: Visual polish issues, like the wrong color shade and cramped chart margins. Treat the AI output as a first draft. Either hand off the file to a designer or use a second prompt to correct specific visual attributes, such as exact color hex, label positions, and margin sizes.
3) Extracting product logic and suggesting an in‑app purchase strategy from source code
What he did: Pointed the agent at a read-only copy of an app’s source tree and asked it to list features, related code paths, and propose an in‑app purchase structure aligned with App Store/Google Play norms.
What the agent produced: A readable product map and a candidate pricing and entitlement plan that accelerated product discussions.
What went wrong and how to fix it: LLMs can misread conditional logic or miss side-effect behavior. Use enforced read-only connections if the platform supports them; if not, host the code in a sandbox and provide view-only tokens. Always require an engineering sign-off on the final mapping.
4) Hardware shortlist for UPS/power needs in a Fab Lab
What he did: Audited device power specs, asked the agent to scan 3D printer requirements and Jackery product pages, and asked for candidate portable power units to replace overloaded UPS devices.
What the agent produced: A rapid shortlist of models and spec comparisons. Jackery ultimately sent four units for lab testing.
What went wrong and how to fix it: Agents can compare specs quickly but can’t validate installation safety. Treat recommendations as a shortlist for vendor or contractor verification, and involve a licensed electrician before any mains or secured-power changes.
5) Financial analysis: vendor spending and loan offer vetting
Vendor audit: Gewirtz asked the agent to scan 12 months of emails and extract bills and receipts from a specific vendor.
“Go through the last 12 months of emails and find all bills or receipts from [my vendor]. Provide me a report on what’s been spent with [my vendor] and for what.”
Result: A task that would have consumed a full day became a roughly 15‑minute run, producing an itemized report that required spot checks rather than full manual reconciliation.
Loan vet: He used a blunt contract prompt:
“Analyze this document very carefully. Look for any indications that it’s fraudulent or, alternatively, any indications that make it clear that it is legitimate.”
The agent did not flag fraud but surfaced deeply unfavorable terms that saved the company from accepting a bad loan. For contract checks, have the agent flag specific red‑flag clauses: unusually high effective APR, prepayment penalties, automatic renewals, assignment clauses, and ambiguous indemnities, and then route the flagged clauses to legal for confirmation.
Concrete example prompt + redacted example output you can reuse
Try this redacted vendor-audit prompt as a starter. Replace bracketed items before running:
“Search my mailbox for messages from [[email protected]] in the last 12 months. Extract invoices, dates, amounts, and a one-line description for each. Produce a CSV with columns: invoice_date, invoice_number, amount_usd, description, message_link.”
Illustrative (redacted) excerpt of expected output:
2025-02-14, INV-1001, 2, 450.00, “Monthly hosting fee”, [link]
2024-11-01, INV-0954, 1, 200.00, “Equipment maintenance”, [link]
Security, privacy, and governance, practical specifics
- Connector scope: Use narrow OAuth scopes, read-only where available. Avoid granting write or admin permissions unless strictly necessary.
- Temporary access: Create a time-bound token or temporary account, run the job, export results, then immediately revoke access. Log the revocation in your admin console.
- Audit trails: Ensure the platform logs which prompts accessed which data. Export that log for internal compliance reviews.
- Data loss prevention (DLP): Integrate DLP scanning on outputs that contain sensitive fields (PII, financial account numbers, health data) and block exports if regulations (GDPR, HIPAA) apply.
- Sandboxing code access: When sharing source code, prefer read-only repository tokens, ephemeral clones in sandboxed environments, or PR-only views that prevent modification.
Prioritized checklist to run your first safe workflow
- Create a narrow-scope account or token (read-only) dedicated to the task.
- Prepare a precise prompt that limits scope (date ranges, sender addresses, file paths).
- Run the job and export results to a secure storage location you control.
- Revoke access immediately and confirm revocation in logs.
- Apply a short human verification checklist before acting on outputs (see starter workflow below).
Starter workflow you can copy this week
Pick email summarization or a vendor-spend audit. Example: vendor audit.
- Prompt (use the redacted prompt above).
- Human checks (3-step): verify top 5 amounts against original attachments; cross-check vendor name against an authoritative ledger; confirm any contract clauses flagged with legal.
- Decide: pay, decline, or escalate based on verified output.
Quick ROI guide: multiply the average hourly rate of the employee doing the task by hours saved per run, then subtract verification time to estimate net savings. If a single run saves 6 hours per month and the effective loaded hourly cost is $60, that’s about $360 per month saved before counting error prevention.
Where agentic AI fits, and where it doesn’t
Use agents for triage, discovery, and first-pass synthesis. They speed up grunt work and surface problems faster. They do not replace domain experts. Always require human sign-off for legal, financial, safety, or regulatory decisions.
Key questions (short, honest answers)
-
Can agentic AI save time on real small‑business tasks?
Yes, individuals have reported drastic reductions (e.g., a dataset-to-slide job in ~13 minutes and a vendor-spend audit that dropped from a day to ~15 minutes). Results will vary by task complexity, prompt quality, and verification needs.
-
Should I give the AI access to email and files?
Only with narrow, time-bound permissions. Use read-only tokens where possible, run a single scoped job, export results, then revoke access and retain the audit logs.
-
Are outputs reliable enough to act on directly?
No. Treat outputs as drafts that need targeted human verification, spot-check names and roles, financial totals, and visual labels before publication or payment.
-
Can agents extract product strategy from source code?
They can accelerate discovery by surfacing features and suggesting pricing frameworks, but engineers and product leads must validate conditional logic and edge cases.
-
Do agentic tools replace accountants, lawyers, or electricians?
No. They help specialists work faster by triaging and flagging issues, but licensed professionals must perform final reviews and assume legal responsibility.
Try one contained workflow this week, an email summary or vendor audit using a read-only token. If the agent frees up an afternoon and helps you catch a costly mistake before it happens, it will have already paid for itself. Keep the human in the loop, log everything, and treat the output as your fast first draft, not the last word.