AI-Powered Digital Co-Workers for Customer Support and Business Process Automation

neople · E-commerce · 2024 · source Agentic: yes · promotional bias: low · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Customer support and business-process automation

Neople needed AI digital co-workers that could remove repetitive customer-service work while preserving human attention for complex cases and making workflow setup accessible to nontechnical teams.

Constraints:

Support workflows contain many customer-specific edge cases
Customers must build trust before moving from suggestions to automation
Quality must hold over time inside each customer's live support operation

Architecture

Neople's Nele digital coworker platform evolved from customer-service response suggestions into automation, action execution, and conversational workflow building. The product connects to customer-specific systems and source material, constrains broad LLM behavior through workflows and guardrails, exposes traces and feedback mechanisms to non-technical customers, and helps users scope repetitive workflows conversationally rather than manually dragging step-by-step automations.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  nele-digital-coworker["nele-digital-coworker"]
  customer-system-connectors["customer-system-connectors"]
  response-suggestion-engine["response-suggestion-engine"]
  automation-and-action-engine["automation-and-action-engine"]
  workflow-guardrails["workflow-guardrails"]
  conversational-workflow-builder["conversational-workflow-builder"]
  workflow-template-library["workflow-template-library"]
  feedback-and-metrics-loop["feedback-and-metrics-loop"]
  nele-digital-coworker --> customer-system-connectors
  nele-digital-coworker --> response-suggestion-engine
  nele-digital-coworker --> automation-and-action-engine
  response-suggestion-engine --> feedback-and-metrics-loop
  automation-and-action-engine --> workflow-guardrails
  conversational-workflow-builder --> workflow-template-library
  feedback-and-metrics-loop --> conversational-workflow-builder

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`nele-digital-coworker`	Customer-branded AI coworker for support and business-process work.	`customer-system-connectors`, `response-suggestion-engine`, `automation-and-action-engine`
`customer-system-connectors`	Integrations into CRM, finance, support, and source-material systems needed for customer-specific work.	—
`response-suggestion-engine`	Human-in-the-loop suggestion path for customer-service agents.	`feedback-and-metrics-loop`
`automation-and-action-engine`	Autonomous response and action-execution layer for repetitive workflows.	`workflow-guardrails`
`workflow-guardrails`	Deterministic and workflow constraints that limit where the agent has freedom.	—
`conversational-workflow-builder`	Non-technical interface where customers describe and scope workflows conversationally.	`workflow-template-library`
`workflow-template-library`	Templates that guide discovery of workflow conditions, inputs, and customization points.	—
`feedback-and-metrics-loop`	Customer feedback, CSAT/closed-ticket signals, first-response metrics, and trace visibility used to improve automation.	`conversational-workflow-builder`

Data flow

A customer defines a Nele and the processes it should support. Early suggestions or automated responses draw on customer systems and source material; feedback buttons and customer-service metrics identify reliable repetitive workflows; deeper automations add action execution; and the conversational workflow builder uses templates and dialogue to scope specific workflows, conditions, and required inputs before deploying them.

3 supporting quotes

“we started off very basic with just suggestions in a contact center.”transcript, stated

“from suggestions went to automation to the actions within customer support and then also to actions and workflows beyond customer support as well.”transcript, stated

“enabling customers to actually teach their people in a conversational way is what really made a difference for us”transcript, stated

Why this architecture

Neople started with suggestions because early LLMs and 2K-token contexts were not reliable enough for direct customer replies. As models improved, the product moved toward deeper automation on narrower repetitive workflows, with feedback loops and conversational workflow definition so non-technical customers could shape automation without becoming workflow-engine experts.

3 supporting quotes

“context windows were, I think 2K at max.”transcript, stated

“Sending out automated answers just wasn't reliable enough.”transcript, stated

“they'd rather have everything regarding orders for example automated than also get suggestions on legal topics for example.”transcript, stated

Models used: LLMs

Scale

Neople had been around for about three years.
“we now have been around for about three years and have 200ish customers”transcript, stated
Neople worked with about 200 customers.
“have 200ish customers that we are working with”transcript, stated
The suggestion phase lasted almost two years.
“the suggestions time frame I think was overall almost 2 years”transcript, stated
Some workflows ran for well over an hour.
“I've seen traces where with workflows that yeah that went on for well over an hour”transcript, stated
Customer first-response rates could move from about 10% to 70%.
“from yeah let's say maybe 10% first response rate to a 70% first response rate”transcript, stated
An ecommerce cancellation workflow may contain at least 40 edge cases.
“there is usually at least 40 different edge cases that you need to take into account”transcript, stated

Alternatives considered (roads not taken)

Automating whole broad processes in one pass

Rejected because: Neople guides customers toward concrete workflows because broad ecommerce processes become huge decision trees.

“If you try to solve all of them at the same time, you're going to end up at least on the back end side of things with a huge decision tree”transcript, stated

Workflow builders requiring technical input syntax

Rejected because: Neople emphasizes conversational workflow definition for non-technical coworkers.

“our focus really is on the conversational aspect as well.”transcript, stated

Suggestions across 100% of contact-center topics

Rejected because: Customers preferred fewer topics automated more deeply over shallow suggestions everywhere.

“tend to prefer to do less topics. So maybe not 100% of the use cases for suggestions in their contact centers but therefore deeper.”transcript, stated

Relying only on vendor-side evaluation

Rejected because: The platform keeps customer feedback mechanisms because customer outcomes are the meaningful evaluation target.

“what are the feedback mechanisms we need to design so that customers are ultimately the eval because that's what matters.”transcript, stated

Patterns

`human-in-the-loop-gate` (automation-ramp)

Neople used suggestion and send-with-final-check flows to let customer-service agents validate repetitive responses before deeper automation.

“one was copy message so that the customer could copy the full message and send that out straight away. And the second one was send straight to customer, which basically gave the agent the ability to just automate it, but having that really tiny final human check”transcript, stated

`data-quality-loop` (customer-feedback-loop)

Customer use of suggestions and feedback controls informed which repetitive workflows were safe to automate.

“customers were obviously giving a lot of feedback and there's also a process in place on those suggestions to give direct feedback. So we had a very good understanding of what's going well, what's not going well.”transcript, stated

`tool-use-orchestration` (business-process-automation)

The agent moves beyond drafting replies by checking order state, changing data in another system, and then answering the customer.

“the human agent would have to go and first check, okay, when was the order placed? Is it still in the place where I can change the address, then probably go to a different system and go ahead and change the address?”transcript, implied

`workflow-decomposition` (workflow-design)

Neople guides customers to automate concrete workflow slices instead of attempting to encode every edge case at once.

“we really look at the concrete workflows, specific workflows with our customers and guide them to start with those rather than trying to automate the whole thing. And that really helps to increase that first time right rates”transcript, stated

`structured-output-contract` (nontechnical-workflow-builder)

The conversational builder collects triggers, conditions, systems, and guardrails so a nontechnical user can assemble a workflow through guided questions.

“it will ask the right questions to really understand okay what system are you using what conditions can I do this by what's my trigger like when do I know that I need to cancel an order”transcript, stated

Tradeoffs

How quickly to move from suggestions to autonomous replies.

Chose automation first on repetitive workflows with human feedback still available over attempting full automation across every support topic immediately.

Rationale: Customers showed high confidence and uptake on repetitive workflows, while complex workflows were less reliable as suggestions.

“It was a very clear distinction between the very complex workflows where usually the suggestions would go semi-well I would say and the repetitive workflows where suggestions were quite yeah confident”transcript, stated

Whether to automate broad support coverage or narrower end-to-end workflows.

Chose fewer topics automated more deeply over suggestions across all contact-center topics.

Rationale: Customers preferred complete automation for well-scoped areas such as orders instead of shallow assistance everywhere.

“customers at this point in time tend to prefer to do less topics. So maybe not 100% of the use cases for suggestions in their contact centers but therefore deeper.”transcript, stated

Gotchas (what bit them)

Trying to automate a whole support process at once creates an unmanageable decision tree.

Root cause: Customer workflows include many edge cases around order state, delivery state, and product conditions.

Resolution: Start with specific concrete workflows and expand iteratively.

“there are so many edge cases to be taken into account. If you try to solve all of them at the same time, you're going to end up at least on the back end side of things with a huge decision tree”transcript, stated

A fifty-page process document may still be too ambiguous for automation.

Root cause: The real business problem and edge cases are often buried inside procedural manuals.

Resolution: Use the conversational workflow builder to decompose the process into triggers, conditions, guardrails, and testable steps.

“they come to us with a 50page manual on what their order cancellation process is. And I have read through many of these documents. And I don't even necessarily always understand what it is exactly that they're trying to achieve”transcript, stated

The agent must avoid using broad model knowledge when it should be constrained to customer-specific source material.

Root cause: LLMs tend to draw on training data even when the workflow should be limited to a source document or customer process.

Resolution: Neople invested in evaluations to make source-bounded behavior reliable.

“LLMs want to use all their training data and you don't want them to use their training data and I've encountered this exact problem of no, you're limited to this source. document. It's really hard”transcript, stated

Evaluation

Neople measures automation quality through customer feedback on suggestions, live customer-support metrics, and downstream customer feedback mechanisms.

Metrics: suggestion acceptance · customer feedback · first response rate · resolution time · first-time-right rate

“these buttons that you added Yob are basically a measurement from the customer. I could use this. I had to edit this, which is really amazing.”transcript, implied

“within CS you often have quite the standard kind of metrics such as first response rates, resolution time, things like that. And so by the doing its job well on those repetitive workflows”transcript, stated

Outcomes

Human agents can focus on complex cases while Nele handles common repetitive requests.
Customers moved from suggestions to automated replies and then to action-taking workflows.
Neople extended the pattern beyond customer support into broader business-process workflows.

The transcript describes customer adoption across repetitive support workflows and expansion into actions beyond support, while the summary reports roughly 200 customers.

“the nele is handling the repetitive like common questions that come up so that the human customer success agent can focus on the more complex cases.”transcript, stated