A South African SME can learn to give longer AI jobs a clear memory, a responsible person and a verifiable stopping point. A week-long quotation follow-up is a useful example: the system must remember what happened, check what changed and know when to hand the work back. Start with those requirements, not an enterprise licence.
The workflow below is a proposed design, not a native Agentforce feature or a tested result. All example records, timings and acceptance rules are hypothetical or proposed. Use them to decide whether your business needs an agent, a scheduled automation or a better manual process.
1. Take the operating lesson, not a purchase instruction
The useful lesson is that a job can continue after the conversation ends, but continuing work needs explicit controls.
In its Source: 14 September 2026 announcement, Salesforce described a long-horizon runtime built around retained context, execution that can resume over time and adjustment through user direction. It named Hunter as the first agent using that runtime. The announcement listed Hunter as a pilot, with general availability planned for November 2026, and described wider use of the runtime as future work.
That is not evidence that every Agentforce agent can already run your quotation process. It also does not establish South African availability, licence eligibility or the cost of your proposed setup. Confirm those points with the vendor if you consider a purchase.
For an SME, translate the announcement into three requirements: preserve progress, resume safely and let the owner change direction. These requirements can guide a custom build without assuming Salesforce access.
2. Define the job before choosing an agent
Define the job as managing follow-up, not winning a sale.
A proposed goal is: follow up on an issued quotation during one working week, record the customer's response and return unresolved matters to the salesperson. The system must not infer that silence means rejection or that a friendly reply means acceptance.
Choose one starting trigger, such as the salesperson marking an issued quotation as ready for follow-up. Exclude draft quotations, disputed quotations and records without a confirmed owner. Record which quotation version the customer received.
Set separate endings:
- A reply needs the salesperson's attention.
- The customer asks for no further contact.
- The proposed follow-up period ends without a reply.
- A record problem blocks safe action.
Only a person should decide whether a reply creates a binding commitment or requires changes to commercial terms.
Before commissioning anything, use the AI agents versus automation comparison. If the job is simply a fixed reminder after a fixed delay, a scheduled workflow may be enough. An agent becomes worth evaluating when varied replies require interpretation and routing.
3. Store a job record that survives a restart
Store progress in a durable business record, not only in chat history.
The proposed record should include a job ID, quotation ID and version, customer record reference, approved contact channel, owner, status, last checked time, next review time and communication references. Keep the quotation in its authoritative system rather than copying an entire customer file into every prompt.
Use a small set of proposed statuses: ready, waiting, review required, blocked and closed. Separate the status from the closure reason. A closed job with no reply is different from a closed job handed to a salesperson.
Source: OpenAI's structured outputs documentation describes schema-constrained responses. That can support consistent fields when extracting a reply classification, but a correctly shaped record is not proof that the classification is true. Handle refusals and incomplete responses as exceptions rather than filling gaps.
For each extracted fact, retain a reference to the source message. The reviewer should be able to inspect the customer's words. The custom AI agent glossary entry provides a shared term for the system being assessed; the job record provides its practical boundaries.
4. Assign responsibility and narrow the available actions
Give the salesperson ownership of the job and give the system only the actions needed to prepare it.
Under this proposed design, the system may read the approved quotation record, inspect the linked message thread, draft a follow-up and create a review task. It may not change pricing, promise delivery, alter bank details, issue a revised quotation or mark a deal as won.
Source: OpenAI's function calling guide explains that the model requests a tool call and the application executes the corresponding code. Use that separation as a control point: application code should verify permissions, record identity and current status before any action runs.
Name a backup reviewer for owner absence. An unattended approval request must not become permission to send.
Ask an appropriate person to assess privacy, security and communication obligations, including the lawful handling of contact information and opt-out requests. Restrict access to the relevant records and agree a retention approach. These design choices require human judgement; a model prompt is not a compliance decision.
5. Schedule checks and make retries safe
Wake the job for a reason, then recheck the record before acting.
A hypothetical schedule could inspect eligible jobs each working morning in South African Standard Time and prepare follow-ups on the second and fourth working days. The business should propose its own calendar, public-holiday handling and contact windows. A week-long job need not mean a model runs continuously for a week.
At each check, read the latest reply, quotation version, owner instructions and communication log. A reply received overnight should cancel a pending generic reminder and move the job to review.
Use a unique action reference combining the job, quotation version and follow-up step. Before sending, check whether that action already has a recorded outcome. After a timeout, inspect the messaging system before retrying: a missing acknowledgement does not prove that nothing was sent.
If connectivity fails, preserve the last confirmed state and mark the action outcome as uncertain. The proposed rule is to block another send until the application or a person resolves that uncertainty. Resume from evidence, not from the model's recollection.
6. Put human approval at the customer-facing boundary
For an initial evaluation, require approval of each outbound follow-up and any consequential record change.
The review screen should show the recipient, channel, quotation version, exact message, latest customer reply and reason for the proposed action. Approval should apply to that specific message and record version, not to an open-ended instruction to keep contacting the customer.
Source: n8n's human-review documentation describes pausing selected tool calls for approval or denial. This is one documented implementation pattern, not a requirement to use n8n or proof that the whole proposed process is safe.
Under the proposed procedure, a denial cancels that action. It does not invite the agent to try another channel. The reviewer records whether to revise, pause or close the job.
If a new reply arrives after approval but before sending, invalidate the pending action and ask for fresh review. Requests involving discounts, contractual meaning, payment instructions or sensitive information go to the appropriate authorised person. The agent may prepare context, but must not settle those matters.
7. Require completion evidence and compare alternatives
Judge the design by whether it closes or hands over correctly, not by how many messages it produces.
A proposed completion record contains the closure reason, supporting message or delivery reference, responsible person and any remaining task. Where no reply arrives, record that the follow-up window ended without a response. Do not call it a lost sale.
Evaluate the same hypothetical cases with a manual process, a scheduled workflow and an agent-assisted workflow. Include a normal reply, no reply, an expired quote, a duplicate record, an owner absence and a send timeout. Ask the reviewer to inspect wrong classifications, duplicate actions, unresolved states and time spent reviewing.
Proposed acceptance gates are no unauthorised customer messages, no repeat send after an uncertain outcome and a visible owner for every blocked job. These are criteria to test, not achieved results. Any serious breach should pause expansion while the team investigates.
The custom-agent workflow resource can support planning. Use the broader AI automation service to consider whether simpler automation fits before specifying an agent.
Reusable requirements brief for a week-long quotation job
Use this brief as the proposed starting specification and replace its example values before commissioning work.
Proposed job brief: quotation follow-up
- Goal: Manage follow-up on an issued quotation; prepare customer responses for a salesperson, not confirm a sale.
- Start: Owner marks a specific quotation version ready and confirms the recipient and permitted channel.
- Duration: Hypothetical one-working-week window using the business's approved SAST calendar.
- Owner: Named salesperson, named backup and route for unresolved reviews.
- State: Job ID, quote ID/version, customer reference, owner, status, next check, latest reply reference and action log.
- Tools: Read quotation and linked messages; draft follow-up; create review task. Sending requires specific human approval.
- Proposed checks: Inspect fresh replies, version changes, contact restrictions and previous action outcomes before each step.
- Proposed exceptions: Missing recipient, ambiguous reply, duplicate job or uncertain send blocks further contact until resolved.
- Proposed limits: No price changes, delivery promises, bank-detail changes, payment actions or legal conclusions.
- Closure evidence: Customer reply reference, recorded stop request, accepted handover or expired follow-up window, plus remaining task and owner.
- Evaluation: Compare manual, scheduled and agent-assisted handling of the same hypothetical records.
- Proposed acceptance: No unauthorised sends; no duplicate sends; restart preserves confirmed state; blocked jobs remain visible.
Worked walkthrough: a normal job and blocked exceptions
A normal job advances on confirmed information; an uncertain job pauses for a person.
Consider hypothetical quotation Q-104, version 2, issued on Monday. The salesperson confirms the recipient and starts the proposed one-working-week follow-up window. On Wednesday, the system finds no reply, drafts a short reminder and requests approval. The salesperson approves the exact message; the application checks again, sends it and stores the message reference.
On Thursday, the customer replies asking whether delivery can happen the following week. The system cancels Friday's pending reminder and creates a review task with the reply attached. The salesperson checks capacity and responds. The follow-up job closes only when the salesperson accepts the handover, with the commercial question still separately visible.
Now consider hypothetical Q-105 with no confirmed recipient and two imported jobs for the same quotation version. The system blocks contact and presents both records to the owner. The owner confirms the correct contact, identifies the canonical job and closes the duplicate with a cross-reference.
If the reply then says only, “That should be fine,” the system must not mark the quotation accepted. It flags the ambiguity. The salesperson reads the thread and, where necessary, asks the customer to clarify. Releasing the block requires a recorded correction, not an assumption.
FAQs about week-long quotation follow-up
Do we need Agentforce to keep a quotation job running for a week?
No. Specify persistent records, scheduled checks, safe action handling and human review first. A scheduled workflow may meet those requirements without an agent. Salesforce's September announcement offers an operating model to learn from, not evidence that an SME needs its platform. If you evaluate Agentforce itself, verify current availability, South African access, plan requirements and integration support before relying on an announced feature.
What happens if the customer replies while a reminder awaits approval?
The proposed rule is to invalidate the pending reminder and route the new reply to the owner. Approval should be tied to the exact message and record version. The application should check again immediately before sending. If something already went out, preserve that evidence and let the salesperson decide whether a correction is needed. Do not let the agent invent an apology or further commitment.
Can the job finish when the customer has not accepted the quotation?
Yes. Finishing follow-up and securing acceptance are different events. The proposed job can close when its window expires, the customer requests no contact or a salesperson accepts a handover. Record the reason and outstanding work. Ambiguous acceptance, changed terms and contractual consequences require authorised human judgement. A completion status should describe the process outcome, not substitute for evidence of a sale.
Turn the brief into a tool decision
Choose the simplest approach that meets the brief and handles its exceptions visibly.
If your business needs help assessing a longer-running quotation workflow, Symaxx's Custom AI agents service is a route to discuss the requirements. Bring a redacted quotation, a typical reply and one difficult exception, then get in touch to explore whether an agent or simpler automation is appropriate.