Yes. Only when low latency changes a commercial outcome—closing a sale, preventing loss, or meeting a legal window—or when delay creates unacceptable risk. Otherwise batch or near-batch automation will usually deliver the same value at lower cost and operational risk.
- Low latency is worth the cost only when it changes a measurable business outcome or prevents loss.
- Operational staff time and incident response dominate lifetime cost more than per-call model fees.
- A pilot without kill criteria, a handover plan and monitoring will fail or be shelved.
- Choose buy, build or consultant based on product leadership, engineering capacity, and ops skills.
- For most 10–200 person companies, batch automation captures most value at far lower cost.
Signs this is happening in your business
- Support reads every incoming ticket during business hours and still misses SLA windows.
- Sales lose qualified leads because no one contacts prospects within minutes.
- Finance spends hours weekly fixing invoice exceptions that could be caught earlier.
- A ‘hero’ undoes agent actions on call when things go wrong.
- Executives arbitrate machine-made decisions because no rollback exists.
- Customer churn spikes from delays between 30 seconds and 10 minutes.
- Regulatory filings run overnight and sometimes miss a deadline.
- Recurring fraud attempts escalate when not stopped within seconds.
- Multiple teams default to ‘real-time’ for every automation request.
If several of these are true at the same time, the problem is usually the process rather than the people running it.
What this is costing you a year
(Team hours per week x hourly cost of employees x 52) = annual payroll loss
- People doing this work4
- Hours each per week6
- Hourly cost of employees$45
- Weeks52
Annual payroll cost of 4 people spending 6 hours a week on repetitive work.
Find the number before you spend anything
Run the formula above on your three most repetitive processes. If any of them clears a year of a salary, that is the one to bring to a call.
Are real-time AI agents worth it for my company?
Real-time agents are worth building only when latency changes a commercial outcome—closing a sale, stopping fraud, meeting a legal window—or when delay creates unacceptable risk. If speed does not alter outcomes, batch or near-batch automation captures most value at much lower cost and operational burden.
Always-on systems add predictable monthly fees and an unpredictable schedule of incidents. You pay for monitoring, model updates and rollback capacity; plan those before you choose speed.
Take four people each doing six hours a week at a loaded cost of $45 an hour: 4 × 6 × $45 × 52 = $56,160 per year spent on repeat work automation can reduce.
When low latency changes the result — the checklist you must run
Low latency matters when faster responses change what happens. If a seconds-to-minutes delay reduces conversion, breaches a legal window, or causes irreversible loss, explore always-on agents; otherwise treat speed as an expensive habit.
- Does a seconds-to-minutes delay reduce conversion or win rate?
- Does the task have a strict external deadline measured in minutes or hours?
- Does a fast response avoid an irreversible cost such as a transfer or inventory loss?
- Is the value per event high enough to justify constant availability?
- Is event frequency steady and high so hourly or daily batches leave many outstanding?
- Can you measure the effect of speed in a short pilot (one to six weeks)?
The F1 analogy: when speed changes the race
Treat latency like qualifying time: sometimes a single brief window decides the result; other times shaving seconds has no effect. If your process has a decisive moment—like a pit stop—invest in always-on systems and the operations to support them. Otherwise, focus on batch improvements.
What does an always-on agent actually cost to run for a 10–200 person firm
Always-on agents cost more than model calls. Expect monthly spend across inference, event routing, hosting, engineering, monitoring and model maintenance. Staffing and incident response typically exceed raw inference fees and drive the recurring budget.
- Inference and model access: $500–$7,000/month depending on model complexity, concurrency and streaming needs.
- Event pipeline and connectivity: $200–$2,000/month for reliable routing, retries and security.
- Hosting and redundancy: $100–$1,500/month for small clusters or managed services.
- Engineering and devops: $3,000–$15,000/month of staff time across build and ongoing changes.
- Monitoring, alerts and incident response: $1,000–$6,000/month for tooling, runbooks and human time.
- Model updates and testing: $500–$4,000/month for evaluation, retraining and validation.
- Vendor licensing and contract management: variable—count legal and vendor time into operations.
A demo that impresses but has no runbook is a liability.
Where the recurring cost really sits: a per-process breakdown
Break the end-to-end into supported steps. Recurring costs align to event handling, live inference, observability and a human rota for exceptions and rollback; each step needs monitoring and budgeted response time.
- Event handling: continuous fees to accept and route events immediately.
- Live inference: higher costs when you need sub-second replies or high concurrency.
- Observability and alerts: tools and people who watch for errors and performance drops.
- Human response and rollback: people on call to stop bad actions and fix data.
- Compliance and audits: time to run checks and store logs for regulation.
How projects actually fail in production — and what surprised me
Pilots often impress, go live, then fail when production diverges from test data and nobody owns fixes. Common failure sequence: mismatched inputs, noisy alerts, damaging actions, cost spikes and alert fatigue. Plan detection, rollback and ownership before launch.
- Mismatch between test data and production; detection: spike in exceptions; mitigation: route suspect items to humans.
- Agent takes a side-effect action (refunds, order changes); detection: customer complaint or reconciliation mismatch; mitigation: immediate rollback and one-click disable.
- Model performance drifts after a product change; detection: rising override rate; mitigation: daily sampling and human review with rollback.
- Alert fatigue makes teams ignore alarms; detection: missed P1 incidents; mitigation: reduce alerts, set clear SLOs and assign on-call ownership.
- Cost spike from unexpected volume; detection: spend above forecast; mitigation: budget caps and throttles that fall back to batch.
Failure mode: hallucinations and actions with side-effects
Agents that can act can be confidently wrong. Action failures are plausible but incorrect steps that create real cost. Prevent them with guardrails, manual approval for high-value moves and an audit trail with immediate rollback.
- Block direct money movement or contract acceptance without human sign-off.
- Run an intent-confidence gate that routes low-confidence cases to humans.
- Keep the agent read-only for the first production quarter on core financial actions.
Detection, SLOs and MTTR you must set before launch
Set SLOs and a target mean time to recovery before you start. Track availability, false-action rate and incident-to-resolution time. Without measurable targets the system becomes a noisy cost centre that surprises finance.
- Availability SLO: set a measurable uptime target and track minutes of downtime per month.
- False-action rate: count actions that required reversals per 1,000 actions.
- MTTR: measure hours from alert to rollback or fix and set a clear target to meet.
Most always-on projects fail because no one budgeted for human on-call time.
Who should own the agent and how the organisation must change
Assign a single accountable owner for the agent’s performance, cost and safety. Ownership depends on purpose: product teams own customer-facing capabilities; platform or ops teams own cross-company automations; bring a consultant if you lack in-house skills.
- If it impacts customers directly, product should own it and prioritise backlog changes.
- If it’s cross-company, a platform or ops owner must run SLOs and vendor relationships.
- If you lack skills to operate it, engage a consultant to scope and set up the first 90 days, then hand over.
Org requirements: skills, runbooks and handover
You need three skills: someone who understands business rules, an engineer who can deploy and patch, and an operator who can respond to incidents. Create a one-page runbook, an alerting policy, an escalation matrix and a rollback switch before launch.
- Runbook: step-by-step for common failures and a single person who can execute a rollback.
- Alerts: limit to P1 and P2 for the first quarter to avoid fatigue.
- Handover: consultants should deliver training sessions and a 30-day shadowing period.
Identify your biggest operational bottleneck
Most owners already know which process is the problem. What they do not have is the annual figure attached to it. That is a 15-minute conversation, not a project.
Buy, build or use a consultant: which path fits your company
Buy when you need speed-to-deployment and accept constrained customization. Build in-house if you have product leadership and engineering capacity. Use a consultant to scope ROI, define SLOs and deliver an operational playbook ready for handover.
- Buy: fastest to deploy but watch for vendor lock-in and limited safety controls.
- Build: highest control and lower long-term cost if you have staff, but slower to reach production.
- Consultant: best for scoping ROI, designing SLOs, and delivering a pilot your team can run.
How to pick the first scope and run a pilot that survives
Choose a narrow, high-frequency, high-value flow you can measure in weeks. Score candidates by Frequency × Value × Latency-sensitivity and pick the top one you can instrument in under four weeks. Require a kill switch, success metrics and a handover plan before you build.
- Pilot length: 4–8 weeks including a one-week baseline measurement.
- Kill criteria: a defined metric drop, a cost threshold, or more than N safety incidents in a week.
- Success metrics: hours saved per week, absolute conversion lift, and reduced override rate.
Speed without a rollback is a live grenade; you need both or nothing.
Design of a pilot that avoids the 'impress and die' fate
Staff the pilot and budget the first 90 days of production. Assign the internal owner, train people who see agent output, and enforce a two-week post-launch check-in cadence so the pilot is not shelved after the first incident.
- Instrument the baseline: measure current time spent and outcomes before changing anything.
- Force a manual override path for the first month and record every override reason.
- Roll out incrementally: start read-only, then escalate the agent's authority after stable performance.
Where batch or near-batch automation is a better choice
For most 10–200 person companies, batch automation captures most value at far lower cost and risk. Choose batch when latency does not change outcomes, volume is bursty, or humans already introduce hours-long delays that make sub-second gains irrelevant.
- Monthly reporting and reconciliations: run overnight when staff are idle.
- Non-urgent notifications and digests: send a single daily summary.
- Processes with human-in-the-loop response times of hours: sub-second timing buys nothing.
- Complex, high-trust operations: keep them manual or batched until you can invest in solid monitoring.
Concrete examples where I would build a real-time agent
I would build always-on agents for fraud prevention where seconds allow loss, for inbound sales contact when minutes materially affect conversion, and for safety or compliance triggers that must fire within minutes. Each requires a committed operations plan and measurable outcomes within weeks.
A practical pilot checklist you can run this week
Run a tight six-step pilot: measure baseline, score flows, write a runbook, implement read-only recommendations, set kill criteria with a budget cap, and hand over to a named owner. Follow the sequence to reach a clear go/no-go.
- Measure the baseline workload and outcome for one week.
- Score potential flows by Frequency × Value × Latency-sensitivity and pick one.
- Write a one-page runbook and identify the on-call person.
- Implement read-only automation first and measure recommendations versus human action.
- Set kill criteria and a budget cap that falls back to batch automatically.
- Handover: train the internal owner and schedule a 30-day operational review.
Next step you can take tomorrow
Run the team-cost arithmetic at the top for your three candidate processes and score them with Frequency × Value × Latency-sensitivity. That single exercise will tell you whether to explore real-time. If you want help mapping that to a pilot and budget, book a scoped assessment: How to work together.
Where I have done this work and how I help
I design pilots, run scoping workshops and hand over production runbooks. I offer advisory and build engagements that cover the business case and the operational plan. See Services for what an engagement includes.
For examples of how this approach plays out, see the Portfolio and case studies. To understand my background and why I hold these positions, read About the author.
Where businesses waste the most time
- Re-keying customer dataWastes staff hours and introduces data entry errorsAdd a direct connection so data entered once flows to the other systemA person still checks: Person verifies exceptions and edits records flagged for low confidence
- Manual invoice matchingFinance spends large blocks of time matching lines and chasing approvalsRun nightly matching and surface only mismatches in a single dashboardA person still checks: Finance reviews and resolves the exception list
- Slow lead follow-upLeads cool off and conversion drops when no one replies within minutesNotify sales immediately for high-value leads and send a daily digest for the restA person still checks: Sales rep contacts high-priority leads within the response window
How this compares with what you already run
| Operational area | The manual way | The automated way | Annual business impact |
|---|---|---|---|
| Reconciliation | People pull reports and match line items daily | Automated nightly matching with exception list | Reduces annual staff hours on matching |
| Customer response | Agents read new tickets and reply during business hours | Batch summaries and responses twice daily | Lowers annual backlog and staffing costs |
| Lead qualification | Sales reps check inbound leads manually | Immediate routing for high-value leads, rest batched | Increases yearly conversions for urgent leads |
| Invoice processing | Finance keys invoices and checks for mismatches | Automated read and match overnight, exceptions flagged | Cuts annual processing hours and approval delays |
| Compliance reporting | Staff assemble monthly reports manually | Automate collection and run checks nightly | Reduces time spent yearly and lowers late filing risk |
| Incident triage | People scan alerts and act in office hours | Real-time triage with human override for high-risk cases | Prevents costly minutes-sensitive incidents each year |
What the end-to-end workflow looks like
- Receive invoiceThe invoice arrives in the shared inbox and is captured automatically.Extracts invoice details into the system
- Match to purchase orderThe system attempts to match the invoice to an existing purchase order.Automatically compares amounts and line items
- Detect exceptionsMismatches or missing approvals are flagged.Runs rules to mark exceptions for review
- Human reviewA finance person resolves the exception or approves the invoice.Stays human
- Post approvalApproved invoices are scheduled for payment and logged.Creates the payment entry and updates records
- Archive and reportInvoice and resolution details are stored and included in monthly reports.Generates the reconciliation report nightly
The metrics that tell you it worked
| Metric | What it tells you |
|---|---|
| Hours saved per week | Total staff hours no longer spent on the task; shows direct labour reduction. |
| Conversion within 5 minutes | Count of conversions within five minutes of first contact; measures latency impact on sales. |
| Override rate | Share of automated decisions humans reverse; measures accuracy and safety. |
| Incidents per month | Number of production incidents requiring human intervention; tracks reliability and burden. |
| Time to rollback (hours) | Average hours from detecting a bad action to full rollback; measures MTTR capability. |
| Percentage of events handled without human touch | Share of events the automation resolves; shows remaining human load. |
When this is the wrong move
- When the event does not change outcomes within minutes or hours and is cheaper to batch.
- When you lack a staffed on-call rota and a clear owner for SLOs and incident response.
- When the value per event is low and volume does not justify always-on costs.
- When data quality is poor and will trigger frequent false actions without cleanup.
- When legal or compliance rules require human approval for the action the agent would take.
The mistakes that cost the most
- Building speed firstTeams prioritise sub-second response and ship without runbooks; in production the system generates noise and the project is disabled.
- No rollbackAn agent makes a damaging action; without an easy rollback it takes hours to contain and privileges are restricted permanently.
- Missing on-call staffingAlerts arrive out of hours, nobody responds, incidents escalate and trust in the system collapses.
- Automating a broken processThe team automates the current manual process instead of fixing it first; the automation scales bad inputs and multiplies work.
Frequently asked questions
Can I start with a batch system and add real-time later?
Yes. Start with a batch or near-batch flow to prove the business case and reduce repetitive work. Use the metrics from that phase to decide if the incremental benefit of always-on justifies the extra operational cost.
How many people do I need on call for an always-on agent?
At minimum, name a single owner and set an on-call rotation covering critical hours. Small firms often use a part-time engineer plus a business owner; larger setups need formal on-call shifts and documented handover.
What is a safe pilot scope for a 60-person company?
Pick one repeated process where speed could change the result—lead triage for high-value prospects or front-line fraud checks. Keep the agent read-only at first and measure lift over four weeks.
How quickly will the costs appear?
Some costs appear immediately: model access and engineering time. Monitoring and on-call costs appear as soon as you move from demo to production; budget the first 90 days explicitly.
Should I hire a consultant to scope this?
If you lack in-house production automation experience, a consultant shortens the learning curve and produces the runbook and kill criteria you need. I scope pilots and deliver operational playbooks as standard services; see Services.
What metrics prove a pilot succeeded?
Clear, measurable improvements such as hours saved per week, a defined conversion lift in absolute numbers, and a reduced override rate. Predefine targets and a measurement period before launch.
Stop paying salary for robotic work
Bring your messiest process to a 15-minute audit and we will map where the hours actually go, live on screen. You leave with the number whether or not you work with me.