Forward-deployed AI pilot

Forward-deployed AI engineering pilot

Salt Technologies embeds an engineer with your team to test one workflow. See a rejected AI draft, cost per accepted task and the client’s rollout decision.

Worked example of a forward-deployed AI pilot

A distributor’s support team wants to spend less time answering “Where is my order?” This fictional four-week pilot shows what a forward-deployed Salt engineer and a client reviewer would test together. The ticket, costs and results below are illustrative.

Download editable text

Before and after the pilot

The support lead selects 100 anonymized historical tickets spanning ordinary shipments, partial orders, missing scans and conflicting records. These cases are held back from development. Reviewers replay each ticket using only the records available when the customer asked. During this evaluation, all replies stay in a test queue.

StepManual baselineAI-assisted workflow
Find the orderAn agent matches the requester and order number, then opens the order system and carrier record.The integration checks the same account and order, then retrieves the timestamped status and carrier events using approved read-only access. A failed match goes to the agent.
Prepare a replyThe agent reconciles the records, writes a reply and decides whether a delivery date is supported.The system drafts a reply with links to the evidence. If the sources are missing or conflict, the intended behavior is to return the task to the agent.
Resolve the ticketThe agent sends the reply and logs the disposition. The baseline averages six minutes per resolved ticket.The reviewer checks the draft against the sources, accepts it, edits wording or rejects it. The log records the decision, reason and time through final handling, including manual fallback.

Salt’s engineer observes the work, connects the approved records, implements fallback and delivers the evaluation log. The client’s support lead owns the definition of a correct reply and authorizes any live release.

A draft the reviewer rejects

Ticket 047 asks whether an order will arrive Friday. The order system says “packed.” The carrier record says “label created” and contains neither a pickup scan nor a delivery estimate.

Review recordWhat happened
Generated draft“Your order has shipped and will arrive on Friday.”
Reviewer decisionReject. A shipping label establishes neither carrier possession nor a Friday delivery. The reviewer records an unsupported delivery promise and takes over the ticket.
Manual replacement“Your order is packed and a shipping label has been created. We’re waiting for the carrier’s first scan. We don’t have a confirmed delivery date yet.”

Ticket 047 counts as a rejection even though the reviewer ultimately resolves it. Its full handling time is included in pilot cost.

What does one accepted task cost?

An accepted task is a draft the reviewer can use as written or after a wording-only edit. A factual rewrite counts as a rejection. In this example the client measures through final resolution, so the rejected drafts and cases that received no draft still incur labor.

Reviewer outcomeTicketsMinutes per ticketTotal minutes
Approved as drafted702140
Approved after a wording edit10440
Rejected and completed manually12896
No draft; completed manually8864

Of 100 tickets, 80 produce accepted drafts. Total handling time is 340 minutes. At $36 per hour, labor costs $204.00. Model and retrieval usage adds $12.00 across the full batch. Cost per accepted task = $216.00 ÷ 80 = $2.70.

How does that compare with the manual queue?

Use all resolved tickets as the common denominator. The manual team spends 600 minutes at $36 per hour, or $360.00 for 100 tickets ($3.60 each). The assisted team resolves the same 100 tickets for $216.00 ($2.16 each), including every manual fallback.

These are variable operating costs. For a rollout budget, add the pilot fee, integration maintenance, monitoring and ongoing evaluation. At this workload the illustrated difference is $1.44 per resolved ticket before those costs. The client would need its own volume and fixed-cost estimate to calculate payback.

Would the client roll this out?

The support lead holds the live rollout. Before testing, the support lead set four gates. All four must pass. The reviewer caught the bad promise before a customer saw it, but producing it still fails the quality gate.

Gate set before evaluationObservedDecision
At least 80% of eligible tickets produce an accepted draft80 of 100 tickets (80%)Pass
Average handling time at most 4 minutes3.4 minutes across all resolved ticketsPass
Variable cost at most $3.00 per accepted task$2.70Pass
Zero unsupported delivery promises in drafts shown to reviewersOne in 100 tickets, on ticket 047Hold

Salt’s engineer changes the shipping-status check so a delivery promise requires a carrier estimate and a label-only record stays pending. The client runs a fresh held-out batch of 100, including label-only cases. If every gate passes, the support lead may authorize 10% of eligible tickets with human approval on every reply. After 50 live tickets the lead reviews errors, cost and time before expanding. A new unsupported promise pauses drafting and returns the queue to manual handling. If the revised pilot still misses a gate, the client keeps the manual process and decides whether another iteration is worth funding.

Plan the evaluation set for your own AI pilot

Discuss a paid pilot for your workflow

Relevant experience

Review the work behind the conversation.

Salt’s clinical intake case documents a guided assessment workflow with professional review. It illustrates how review fits into software. The four-week pilot below is a proposed engagement with its own success criteria.

Where we can help

Which workflow fits an embedded engineer?

These are starting scopes. Your systems, decision owners, and delivery constraints shape the final proposal.

Customer implementation

For a software vendor, the engineer works with your implementation lead and customer users on one integration. Agree access, escalation and who owns the code before the first session.

See software engineering services

Operations workflow

For example, draft a reply to an order-status request from approved order records. A reviewer checks the draft before sending. Missing or conflicting records should send the task back to the team.

Explore AI and automation

Deployment and handover

Identify who maintains the integration, where traces may be stored, how to switch off the AI step and how staff resume the existing workflow.

Explore deployment and modernization

From discussion to delivery

What happens in a four-week AI pilot?

Explore our SPARK framework
  1. 01

    Week 1 · Observe and measure

    Watch a user complete the task. Agree the sample, baseline handling time, unacceptable errors, data permissions and a written decision threshold.

  2. 02

    Week 2 · Build a narrow integration

    Connect approved sources and prototype one path. Keep a set of representative cases out of development for the final evaluation.

  3. 03

    Week 3 · Test with reviewers

    Run the held-out cases. Record acceptance, edits, unsupported answers, abstentions, elapsed reviewer time and cost per accepted task.

  4. 04

    Week 4 · Make the release decision

    Review failures with the workflow owner. Deliver the evaluation, code and operating notes. Choose a limited rollout, a specified revision or a stop, with a named owner for each next action.

Delivery ownership

Who works with the engineer?

Embedded with your operations team

Your process owner supplies examples and reviewer time. Salt scopes the integration and evidence needed for a decision on one workflow.

Discuss an operations workflow

Embedded with a software vendor

Your product team owns the roadmap. Salt works alongside implementation staff to solve a customer-specific deployment and document what can be reused.

Explore delivery partnerships

Working with your US team

Agree time-zone overlap, communication channels, decision-makers, repository access, and review cadence before delivery begins. Scope, staffing, commercial terms, ownership, and post-launch responsibilities belong in the proposal and agreement.

Before you begin

Questions to resolve
before committing.

A forward-deployed engineer works directly with the people using or deploying a product. In Salt’s AI pilot, that engineer observes the task, connects the systems, tests with users and hands the result to its ongoing owner.

Four weeks is a proposed schedule for one paid pilot. We quote after reviewing the workflow, access, staffing and evaluation needs. Procurement or data-access dependencies can change the schedule.

We agree working hours, access to users and remote collaboration with your team. Any on-site requirement must be discussed and agreed in the scope.

You receive the results, examples of failures and a recommendation. The workflow owner can keep the existing process, commission a targeted revision or approve a limited rollout under defined controls.

Start with a short brief

Discuss a paid AI pilot.

Tell us who performs the task today, which systems it touches and how you would judge a better outcome. An anonymized example is enough to start.

Discuss a paid AI pilot