Abstract
Mass-market SaaS products model the process their customers have in common. Where an organization's process differs, the difference is absorbed by manual work: side spreadsheets, duplicate entry and email. This paper describes that gap in technical terms and sets out the approach LekkerLogic uses to close it: a custom application on a shared data layer, integration APIs between the systems already in place, AI assigned to defined workflow steps under human approval, add-in modules for Microsoft Outlook, and a per-task model for measuring return.
30-second read
- §1SaaS is built for the many, with no direct connection to your business. What only you know about your process ends up lost in SaaS solution gaps and never truly addressed in the operation.
- §2–3A custom build, designed with the people who do the work, models your own data and connects CRM, ERP, mail and outside data through one data layer.
- §4AI is assigned to defined steps: the model interprets, code verifies, a person approves, and only then is a record written.
- §5Outlook add-ins put those actions in the inbox, on the same API and access rules.
- §6–7Return is calculated per task from measured inputs and realized through adoption. If a mass-market product already fits, buy it.
1The coverage gap
A SaaS vendor serves thousands of organizations with one data model and one set of workflows. That is efficient for functions that are the same everywhere, such as payroll, accounting and email, and those should be bought, not built.
For processes specific to an organization (how quotes are built, how orders arrive, how ship dates are committed, what customers are allowed to see) the product covers part of the requirement. The remainder appears as four recurring symptoms, each with a structural cause.
- A spreadsheet kept beside the system of recordThe product's data model has no place for entities the process depends on.
- The same data keyed into two or three systemsThe systems share no keys and no integration.
- Paid seats for people who only readAccess is tied to licensing, not to role.
- A requirement waiting on the vendor's roadmapThere is no extension point where the process differs.
1.1What a vendor cannot know
A mass-market product is designed for the average of its customer base, and its vendor has no direct connection to any one business. It cannot know what only that business knows: which customers get which terms, why an order is held, what a committed date really depends on, which exceptions are routine. That knowledge lives with the people doing the work, not in any product specification, so the product cannot encode it. It can offer configuration, and the business supplies the rest by hand. A custom build starts from that knowledge instead (Section 2).
Table 1Comparison by technical dimension.
| Dimension | Mass-market SaaS | Custom build |
|---|---|---|
| Data model | Fixed entities, with custom fields inside vendor limits. | A schema designed around the organization's own entities and terms. |
| Workflow logic | Configurable within the product's states and rules. | Written to match the actual process, including its exceptions. |
| Integration | Vendor-provided connectors, within API limits. | Direct integration with CRM, ERP, mail and external APIs. |
| AI | A general assistant with limited reach into business data. | Models assigned to defined steps, with verification and approval. |
| Access control | Tied to per-seat licensing. | Scoped by role and company. External users need no paid seat. |
| Change control | The vendor's roadmap and release schedule. | The organization's own priorities. |
| Cost driver | Seats per year. | Build and support effort. |
Custom software carries its own costs: it has to be specified, built and maintained. Section 7 gives criteria for when that trade is worthwhile.
2Method: design with the operating team
Requirements are taken from the people who perform the work (sales, project management, customer service), because the specifics that matter are known only to them. They do not come from a feature list. One small team carries a project through four stages, and the operators review working software throughout the build.
Table 2Stages, activities and outputs.
| Stage | Activities | Output |
|---|---|---|
| Consult | Walk the workflow with its operators. Inventory tasks, systems and data sources. | An agreed definition of "solved" and a task baseline. |
| Plan | Scope, data model, integration map, priorities and timeline. | A plan, with the measure the build will be held to. |
| Build | Short cycles. Working software reviewed by its users. Integration against the real systems. | Software the team has already steered. |
| Deliver | Launch, documentation, training and ongoing support. | A system in production and a measured result. |
Table 3Workflow changes observed in client builds. Clients are not named.
| Workflow | Before | After |
|---|---|---|
| Project delivery for custom-manufactured work | Tracked on a general-purpose board. Purchase orders, lead times and ship dates lived in workarounds, and status was gathered by phone. | Those entities are part of the data model. One view ranks projects needing attention, committed dates are retained when they slip, and suppliers read status through scoped links. |
| Customer parts and quote requests | Stock, pricing, drawings and quotes were requested by phone or email and answered one at a time by sales. | Customers self-serve against company-specific pricing and live stock, and submit a complete parts list in one request. |
| Meeting records | Notes depended on the individual. Third-party note-taking bots were not acceptable on client calls. | A recording uploaded afterward yields notes in a fixed structure, action items with owner and date, and a searchable archive. |
3Reference architecture
Existing systems stay in place. A custom application sits between them on a shared data layer, reached through integration APIs. Data is entered once, in the system that owns it, and appears wherever it is needed.
3.1Data layer
One relational schema (MySQL or SQL Server) holds the organization's entities, with keys shared across the systems it connects. Each entity has a designated system of record. Access is scoped by role and, for external users, by company, so a customer or supplier account sees only its own records.
- Systems of record
- Outlook and email
- CRM
- ERP and orders
- Spreadsheets
- External data APIs
- Consumers
- Web app and dashboards
- Customer and partner portals
- AI tools and agents
- Outlook add-ins
3.2Integration APIs
Table 4Integration patterns, internal and external.
| Direction | Example | Implementation |
|---|---|---|
| Inbound | A won opportunity in the CRM (Salesforce, Microsoft Dynamics) creates a project or order. | REST calls or webhooks, matched on shared keys. |
| Outbound | Ship dates and order data are supplied to the ERP. | A data API that the ERP or its middleware reads. |
| External | Third-party data services. | Server-side only. OAuth 2.0 credentials stay on the server, tokens and responses are cached and shared across sessions, rate limits are honored, and the last good data is served during an outage. |
| Messages and attachments in Outlook. | Add-in modules calling the same API (Section 5). |
4AI in the workflow
Many organizations are uncertain how to deploy AI: which data is safe to connect, where the return is, and who would build it. A general chat assistant does not resolve this. It sits beside the work, a person moves data in and out by hand, and the saving is small and hard to measure. The alternative is to assign a model to a defined step with a known input and output, and to surround it with ordinary software controls.
- 1. InputEmail, PDF, photo
- 2. InterpretAI model
- 3. VerifyDeterministic code
- 4. ApprovePerson
- 5. CommitCRM, ERP, SQL
Input arrives in its natural form: an email, a PDF purchase order, a photo, a transcript. The model converts it to structured data against a fixed schema. Deterministic code then validates the result by matching it against customer and catalog records, checking the arithmetic and flagging anything it cannot find. A person reviews the result beside its source and approves it. Only then is a record written.
The model interprets. It does not decide, and it cannot introduce a customer or part that does not exist in the reference data.
// Stage 2 output after stage 3 matching (illustrative) { "customer": { "name": "Acme Controls", "matched_id": 1042 }, "po_number": "48213", "lines": [ { "part": "ENC-2420-08", "qty": 6, "match": "exact" }, { "part": "ENC-3024", "qty": 2, "match": "none" } ], "status": "needs_review" // line 2 not in catalog }
Table 5Controls applied to AI features.
| Control | Implementation |
|---|---|
| Data handling | Data reaches models only through business APIs that do not train on it, limited to what the task needs and subject to the requesting user's access rules. Data flows are documented for IT and legal review. |
| Human approval | Nothing changes a customer record, a price or an order until a person approves it, with the source in view. |
| Model selection | Set in configuration per task: stronger models for analysis, faster low-cost models for bulk reading. Changing model or vendor is a settings change, not a rebuild. |
| Cost | Results are cached until the underlying data changes, so the same analysis is not paid for twice. |
| Exceptions | Unmatched items go to review and are logged, so reference data improves from real use. AI on public forms is rate-limited. |
5Outlook add-in modules
A large share of business process starts in email. An Outlook add-in is a web application that Outlook loads in a task pane beside the open message. It is declared by a manifest, deployed centrally by the Microsoft 365 administrator, and uses the Office JavaScript API to read the message's sender, subject, body and attachments.
The add-in calls the same integration API as the main application (Section 3.2), so it follows the same access rules and writes to the same records.
- The user opens a message and the add-in reads its context.
- The server matches the sender and any references in the message to an account, project or order.
- The user chooses an action: save the message and attachments to the record, create a task, log the conversation to the CRM, or have AI summarize a long thread.
- The record is written and linked. Nothing is re-typed and the user does not leave Outlook.
6Measuring return
Return is modeled per task, not per project. For each task in a workflow, five inputs are recorded with the operating team during Consult.
- N
- People who perform the task
- F
- Times each person performs it per week
- T0
- Minutes per run today
- T1
- Minutes per run once automated (review and approve)
- C
- Loaded cost per hour
hours_per_week = N × F × (T0 − T1) ÷ 60 hours_per_year = hours_per_week × 48 // working weeks value_per_year = hours_per_year × C
T0 is measured before the build. T1 is estimated in Plan and measured again in production, which gives the figure the build is held to. The model counts labor time only. Error rates and turnaround time are tracked separately where they matter.
Example 1Worked example with illustrative inputs, not a client result. Edit the inputs to model a task of your own.
- hours_per_week
- 12
- hours_per_year
- 576
- value_per_year
- $25,920
6.1Adoption is the multiplier
The model assumes the software is used. N and F are adoption figures: if half the team keeps working around a system, half the projected return never appears. Software that is well built and beautifully executed (fast, clear, and shaped to the task in the operators' own terms) is software people choose to use. That raises adoption, and adoption is what turns a projected return into a realized one. Execution quality is therefore part of the ROI case, not a finish applied afterward.
7When custom is warranted
Custom is not the default. If a mass-market product fits the process, it is the lower-cost choice. A custom build is warranted when several of the following hold.
- aThe process that differentiates the organization is the one its software models worst.
- bStaff maintain parallel records, or re-enter data between systems.
- cSystems of record are not integrated, and people perform the integration.
- dAI is expected to perform workflow steps, and data handling has to be known and documented.
- eLicense cost is driven by users who only read.
Can your systems answer that? If the software you run today can't show strong adoption and a measurable return, let's talk. LekkerLogic is a small team that designs and builds the systems described in this paper: custom web applications, integrations and AI tools.
Let's talk Let's talk: lekkerlogic.com